Symposium F.MT04—Using Machine Learning and Multiscale Modeling to Study Soft Materials and Interfaces

The advancement in modern electronics has led to miniaturization of device and increase in demand for ultra-small and highly integrated device technology. Extreme ultraviolet lithography (EUV) has been utilized to meet the demand for patterning nano-elements of 20nm or less, however has issues regarding stochastic effects causing defects in the system. To resolve this issues, directed self-assembly of block copolymers has been considered as a promising candidate to aid EUV to develop of sub-10 nm patterns. To achieve the sub-10 nm pattern, block copolymers with high Flory-Huggins parameter (χ) is generally required for high quality pattern resolution. However, a few issues arise when using high χ block copolymers. First, high χ values generally lead to a large surface energy difference between the individual polymer blocks preventing formation of perpendicularly oriented lamellar patterns and reduce horizontal alignment of polymer blocks. Neutral brush layer coating on the substrate is an option to reduce the energy difference, but does not fully resolve the issue as it not only requires additional processes and cost but also cannot be applied on substrates with prefabricated patterns. Secondly, high χ block copolymer has slower self-assembly kinetics, leading to increased defect occurrence and self-assembly time. Therefore, a deliberate design for block copolymers with moderately high χ values as well as rapid kinetics is required for facile DSA of perpendicular patterns.
Here, we suggest a newly designed block copolymer that can achieve high χ value while forming perpendicular lamellar pattern. The key to the block copolymer design is incorporating a gradient copolymer block with Si-containing block to promote universal perpendicular alignment of lamellar morphology without the need for additional surface treatment or neutral brush layer. The block copolymer was synthesized by reversible addition-fragmentation chain-transfer polymerization to form a styrene (S)/ 2,3,4,5,6-pentafluorostyrene (PFS) gradient with a Si-containing 4-(tert-butyldimethylsiloxy)-styrene (4BDSS) copolymer block (P(S-g-PFS)-b-P4BDSS). The styrene/ pentafluorostyrene gradient copolymer consists of gradual composition change from the block junction region to the tail, which creates surface energy difference enough to energetically drive perpendicular lamellae formation, well supported by a thermodynamic model. Furthermore, by mixing P4BDSS homopolymer block with the block copolymer, the kinetic factor can be increased to be greater than that of existing high χ block copolymers to accelerate the self-assembly time and reduce the number of defect occurrence.

Atomistic simulation techniques such as molecular dynamics (MD) provide an accurate and precise description of atomic motion, molecular structure and permit the prediction of the physical and chemical properties of molecular system. However, MD requires expensive computation of energy and force which leads to significant computational effort. This severely limits MD applications to biological system and soft matter physics on long time scales.
The usual MD time step is approximately 1/10 of the fastest frequency of the molecular system. In order to accelerate the MD computation, we propose a machine learning approach to propagate the molecular system instead of the usual MD time step. As a first step we developed a machine learning (ML) propagator for hard-sphere systems that propagates the molecular system with each atomic collision as a new time step. The proposed algorithm learns the time evolution of the atomic motion and the collision between atoms, such that the neural network are able to predict the system trajectory, identify the collided atomic pairs and correct the trajectory of the collided pairs for each collision time step. We will discuss the perspective of this newly ML propagator for the acceleration of MD simulations and further application to the molecular system with long time scales.

Thermoplastic elastomer (TPE) is a typical industrial product, in which the microphase separation of block copolymer is utilized. The products can be seen in daily life, elastic fiber, film and adhesive for examples. The dynamic properties such as non-linear stress-strain (S-S) behavior are key issue to design such industrial products. The phase separated structure as well as a polymer chain structure affects the S-S behaviors. However, it is often the case that complicate metastable structures are observed in industrial products, and it is not simple to find the relation between such phase separated structures and S-S behaviors. To tackle the problem, we applied hierarchical simulation and deep learning technique. We chose ABA type triblock copolymers, where A blocks and B blocks form glassy domain and rubbery domain respectively, as a TPE model. Stress-strain curve of wide variety of volume fraction and phase separated structures were investigated by hierarchical simulation of self-consistent field theory and coarse-grained molecular dynamics. Furthermore, we applied 3D-convolutional neural network (3D-CNN) to make regression between the phase separated structures and S-S curves obtained by the coarse-grained molecular dynamics simulation. The predicted S-S curve of untrained structures using trained 3D-CNN showed good agreement with simulation, and high-throughput prediction could be realized comparing to the computationally intensive simulation.

This work was supported by a JSPS Grant-in-Aid for Scientific Research on Innovative Areas: “Discrete Geometric Analysis for Materials Design” [Grant Number 17H06464].

Objectives: The SARS-CoV-2 spike glycoprotein (S-protein), the presumed key receptor for binding to various material substrates, promotes the entry to host cells, and thereby causes infection. The binding dynamics of the S-protein to various substrates including organic, inorganic materials, and living cells is of great interest in biological science and clinical medicine. The all-atomic molecular dynamics (MD) simulations that are known to generate reliable results while mitigating potential biological risks for laboratory research are prohibitively expensive in computing time and resources.
Methods: We develop a coarse-grained (CG) model of the S-protein by intelligently learning the parameters from the corroborated in vitro and in silico data using DNNs on IBM’s supercomputer. The CG parameters are optimized with the protein in water to provide a medium that offered fluidity in vivo. The MD data produced by GROMACS, are fed to the DNN to predict the force field among the CG particles to be simulated in LAMMPS. The loss function that measures the discrepancy between MD values, NN prediction, and CGMD functions, is minimized during the learning process. CG models speed up the all-atomic model by coalescing atoms to form macromolecules and thereby forgoing the unnecessary calculations of the internal degrees of freedom. The CG force field is electro-free and the solvent is modeled using DPD at the mesoscopic scales. The interactions between the proteins, their surrounding solvent, and the surface they bind to are described by formulating a hybrid of the Morse-repulsive and the dissipative terms. Besides, AI is also used to predict the associated temporal scale parameters.
Results: The VMD model builder is used to construct two different CG models, shape-based coarse-graining (SBCG) and residue-based coarse-graining (RBCG), to the S-protein atomic data (PDB: 6VXX) which has 45,246 atoms, and evaluated their robustness with comparison to atomic-scale simulation. The nonbonded Lennard-Jones and bonded interaction parameters in the SBCG model, consisting of 355 beads with approximately 150 atoms per bead, were learned from the all-atom simulations. The bonded interaction parameters were tuned until the stiffness constants deviated less than 25% from those derived from the atomic-scale simulations. The RBCG model consists of 6498 beads, with approximately 7 atoms per bead, built with the RBCG 2007 topology file inside the Martini force field. We then solvated the model with Martini water and ionized the model with Na⁺ and Cl^- ions. After minimizing the CG system potentials, we successfully ran GPU-accelerated simulations with multiple runs to assess the impacts of the incrementally increasing timesteps to 14 fs. For verification and validation, the preliminary RMSD values were compared with the all-atomic simulations as the control, involving a single S-protein solvated in a water box with neutralizing ions at a temperature of 310K. The RMSD trends were converged which means our CG models have stable structures. Minimizing the differences of RMSD values by AI enables us to obtain more accurate parameters for longer simulations.
Discussions and Future Work: Our CG model uses learned parameters from big data to accelerate the conventional all-atomic MD by 4~5 orders of magnitude, without significant loss of accuracy. As the first AI-guided multiscale CG model, it may help enable a thorough investigation of the binding dynamics of the S-protein at varied ambient conditions including temperature and pH values. The limitation of this work may be that depending on different research needs, this CG model could achieve protein chain level accuracy, but not the atomic level. Our AI-guided modeling methodology can be generalized conveniently to study a large collection of materials, e.g., fibrinogen.
Acknowledgment: The project is supported by OVPR&IEDM COVID-19 Grant, SUNY-IBM Consortium Award and Garcia High School Program.

Controlled degradation of micro- and nanogels finds its use in several applications. For instance, in the healthcare field, drugs containing gel particles are utilized for drug delivery. Drug release from such gels can be controlled via degradation of the hydrogel network triggered by various external stimuli. A better understanding of the gel degradation process will aid in the development of such applications. Computational modeling of gel degradation allows one to capture the dynamics of the degradation process depending on the degradation rate constant, initial crosslink density, size of the nanogel particle, and polymer-solvent affinity. To this end, we have developed a Dissipative Particle Dynamic (DPD) approach to capture the degradation of tetra-arm polyethylene glycol based nanogels. These networks are of interest since they have near-ideal network structure thereby exhibiting superior mechanical strength prior to the controlled degradation. Within our simulation approach, we model bond breaking during degradation as a stochastic process. Our results allow us to identify two distinct stages of the nanogel degradation. During the first stage, the bonds breaking within the gel results in the decrease in the crosslink density within the bulk of the gel particle and small tetra-arm pieces breaking off the surface of the nanogel (surface erosion). During the second stage, bulk erosion and reverse gelation occurs. We characterize the transition between the two stages by focusing on the time evolution of the radius of gyration of the largest cluster and reduced degree of polymerization during the degradation process. We show that during the surface erosion, the radius of gyration of the largest cluster increases due to the effective decrease in the crosslink density being a major process, while during the second stage, this trend is reversed. We identify the effects of the gel size and an initial crosslink density on the crossover between the surface and bulk erosion. Our results represent a first mesoscale simulation study of a crossover between the surface and bulk erosion in gels.

Using solvent-free solid polymer electrolytes in batteries would reduce concerns about flammability and potentially decrease weight and cost, especially if their ion conductivity can be improved. Strategies to improve ion conduction in ion containing polymer materials include using bulky anions or using a high dielectric constant polymer. These both effectively decrease the strength of Coulomb interactions between ions at contact and can reduce ion agglomeration, which can improve conductivity. However, strong ion-polymer interactions can slow ion motion, and size asymmetry between ions may increase preferential solvation of cations versus the larger anions, lowering the transference number t₊ (fraction of conductivity contributed by the cation). We study these effects using coarse-grained molecular dynamics (MD) simulations which include a 1/r⁴ potential to capture size-dependent ion-monomer and ion-ion solvation effects. This is the same form as the interaction between an ion and an induced dipole. For block copolymer systems, we have also been guided by theoretical calculations to show the effects of macroscopic phase behavior. From MD, we calculate conductivity from ion mobilities in an external electric field, which improves accuracy versus the typical use of fluctuation dissipation relationships. We assess the impacts of ion size and polymer dielectric strength, including polymer dielectric strength contrast for block copolymer systems, on ion correlations, diffusion, and conductivity.

Proteins are known to undergo a dynamical transition at low temperature, typically between 220 and 240 K. This transition consists in a steep enhancement of the protein atomic mobility which is related to the onset of anharmonic motions. At the same temperature proteins start to be biologically active and thus a full understanding of the molecular origin of this phenomenon is of particular relevance. While the protein dynamical transition always occurs in aqueous environment, with the protein confinement preventing the onset of ice crystallization, the role of water is still a controversial issue.
Here we focus on the low temperature behavior of a non-biological macromolecule, i.e. poly(N-isopropylacrylamide), commonly known as PNIPAM. This polymer is mostly exploited for its thermo-responsive nature, but it is also of interest because of its affinity with proteins in terms of energy landscape and amphiphilic chemical composition. Depending on the specific synthesis protocol, different PNIPAM based molecular architecture can be originated, such as linear chains or microgels, i.e. polymer network particles with a colloidal size. In this talk I will discuss two studies on the low temperature behavior of this synthetic soft material where atomistic molecular dynamics simulations are combined to elastic incoherent neutron scattering experiments.
In the first part I will provide a description of the evidences of the occurrence of a low temperature dynamical transition in PNIPAM microgels, akin to that observed in proteins [1,2]. This study is based on a nanoscale model of a microgel network in water, which quantitatively reproduces neutron scattering experiments. By correlating the information extracted from the analysis of the polymer relaxations times, water self-diffusion coefficients and hydrogen bonding interactions I will show that water-polymer coupling plays a driving role in the phenomenon.
In the second part, I will report the observation of a low temperature dynamical transition also in PNIPAM linear chains, which suggest a wide generality of the phenomenon, independently on the macromolecular architecture [3]. I will also show that simulations allow to identify the specific motions underlying the onset of the transition.

References
[1] Zanatta M. et al. “Evidence of a low-temperature dynamical transition in concentrated microgels” Science advances 2018, 4, eaat5895.
[2] Tavagnacco L. et al. “Water-polymer coupling induces a dynamical transition in microgels” The journal of physical chemistry letters, 2019, 10, 870-876.
[3] Tavagnacco L. et al. “The dynamical transition of hydrated polymer chains” in preparation.

Over the last few decades, a great deal of attention has been given to topologically linked ring polymers by both the physics and chemistry communities. However, the dynamics of such systems remain largely unstudied. Indeed, these systems represent a major theoretical challenge as they are typically subject to both topological and hydrodynamic interactions, which make the equations of motion intractable. We circumvent this difficulty by considering the dynamics of linked rings with pre-averaged hydrodynamic interactions in the spirit of the Zimm model and demonstrate that the symmetries of the rings lead to a hydrodynamic decoupling of the internal ring relaxations. In other words, the hydrodynamic interactions occur only at the ring center-of-mass level. This decoupling dramatically simplifies the problem and actually holds for any averaging procedure, even for non-Gaussian polymers and non-equilibrium conditions. We use this result to design efficient algorithms for Brownian dynamics simulations which offer enormous computational speed-ups and use these methods to study the dynamics of catenated polymers.

Polyelectrolyte (PE) brushes are formed when charge-bearing polymer molecules are grafted to a substrate in close vicinity of each other. PE brushes have been at the forefront of soft materials research for quite some time due to a plethora of applications ranging from ion sensing, emulsion stabilization, nanofluidic energy conversion to drug delivery and oil recovery. The most intriguing properties of PE brushes stem from their responsiveness to external stimuli such as the solvent quality, pH and bulk salt concentration. We use an all-atom Molecular Dynamics (MD) framework to study the effects of varying charge density on the microstructure of densely grafted planar polyacrylic acid (PAA) brushes. Such a variation in the PE charge density due to changes in the degree of ionization of the weakly acidic/basic PE functional groups is typically bought by a variation in the pH of the surrounding medium. We investigate the properties of the PE molecules such as brush height and chain mobility along with the local structure and distribution of the counterions and water molecules within the brushes. Our results uncover several non-intuitive phenomena such as a grafting density dependent non-monotonic variation in the static dielectric constant of water molecules within the brushes with the degree of ionization. Moreover, we observe a remarkable increase in the species storage capacity of the brushes with the degree of ionization as characterized by an enhancement in the number density of both counterions and water molecules trapped within the PE layer. Our results shall help answer fundamental questions regarding the behaviour of weak PE brushes which is vital to improve the efficiency of several PE brush-based devices such as nanomechanical gates, nanofluidic diodes, current rectifiers, etc.

Polyelectrolyte (PE) brush grafted systems have been used in a plethora of applications for their responsiveness to different stimuli. Here, we have used all-atom molecular dynamics (MD) calculations to study PE brush grafted nanochannels with added salt and their response to external electric fields generating electroosmotic (EOS) transport. We have discovered that nanoconfinement creates an overscreening effect inside PE brush layer, which implies a presence of excessive counterions (Na⁺) inside negatively PE brush layer. As a result, in low electric fields the overall EOS flow of the solvent water is dictated by the motion of the excess coions present in the brush-free bulk. Along with a higher concentration of coions in the bulk, higher velocity of coions than counterions in these applied electric fields also helps the solvent to move with coions. Therefore, we find a fascinating scenario where the EOS flow is coion-dictated, while EOS flow has always been known to be counterion-dictated. On the other hand, for a larger electric field, the PE brush height reduces significantly and excludes some overscreening-inducing counterions into the brush-free bulk, which ensures parity in the number of coions and counterions in the bulk. Similar concentration and velocity of coions and counterions in the bulk in this electric field are overwhelmed by the larger residence time of water molecules inside the first solvation shell of counterions which ultimately results in a counterion dominated solvent flow in the direction of the applied electric field. Therefore, we encounter another highly surprising situation where the direction of the nanochannel EOS transport could be switche by merely changing the magnitude of the applied electric field.

Acknowledgement: This work has been supported by the U.S. Department of Energy Office of Science grant DE-SC0017741.

Multiphase polymer systems, including polymer nanocomposites (PNC), have a wide range of current and potential uses due to extraordinary tailorability of a variety of physical properties with small composition or processing changes. However, the details of the polymer physics controlling these changes is poorly understood. Modeling the interphase region in PNCs is crucial for achieving both better understanding of underlying physics and rational design of functional PNCs. Properties of the interphase region depend on that of the polymer matrix, nanoscale reinforcements, and surface chemistry in addition to microstructural characteristics of the distribution of nanoparticles in the PNCs. However, identifying the local property profile and the length scale of the interphase remains challenging. In this work, we establish a multiscale modeling framework to study the role that microstructure plays in determining bulk mechanical and viscoelastic properties of PNCs, bridging simulations at molecular and continuum level. Sensitivity of the structure--property relationship is also discussed.

Liquid and ion transport in nanofluidic systems have wide applications in ionic sensing, energy generation, bio-medicine, and the development of biomimetic systems. In this study, we analyze liquid transport by applying thermal gradient in a nanochannel functionalized with backbone charged pH-responsive polyelectrolytes (PE) brushes which is modeled using recently developed Augmented Strong Stretching theory (SST). We find that grafting nanochannels with charged polyelectrolyte brushes leads to enhanced thermoosmotic (TOS) fluid flow when compared to the bare nanochannels, as opposed to the general understanding that the polymer/PE brushes retard the flow owing to increase in drag due to the brushes. This is due to the fact that the PE brush induced localization of the Electric Double Layer (EDL) charge density away from the nanochannel walls results in a larger background flow for the migration of these ions. We quantify the electric field induced due to TOS flow and analyze its three components: ionic, osmotic, and thermal components. For various chosen parameters, we witness that the TOS velocity is greater in magnitude for the nanochannels grafted with backbone charged PE brushes when compared to the bare nanochannels or the nanochannels grafted with PE brushes with charge only at the end away from the wall.

There are few researches for constitutive modeling of shear thickening materials, though it is essential to predict their shear stiffening behavior, understand the mechanism of rate-dependent stiffness, and design new absorption devices. In this study, the constitutive model of shear thickening gels (STG) was developed by assuming STG consisting of soft and hard segments, the volume fraction of which are dependent on the shear rates. As such, two different rheological models were used for the soft and hard segments. The parameters of each rheological model were determined by carrying out the lap shear and pulling-out tests of STG. Finally, the developed constitutive model was validated by simulating the impact behavior of multi-layered sandwich composites including STG and polymer films and comparing the results with experiments.

The need to develop environmentally friendly and energy efficient technology for gas separations can be fulfilled by polymer Mixed Matrix Membranes (MMMs) with Metal organic frameworks (MOFs). Gas separations via selective transport in MMMs have shown to enhance selectivity and permeability. In these MMMs, polymers, MOFs and the polymer-MOF interface are known to play an important role in determining the gas separation mechanism. However, fundamental understanding of the parameters determining the compatibility of the polymer-MOF interface and its effect on the separation mechanism is yet to be developed due to limitations of experimental characterization methods. In this study, we have performed all-atom molecular dynamics (MD) simulation of a model polymer-MOF MMMs of polystyrene (PS) and polyacrylic acids (PAA) polymers with IRMOF-1 and HKUST as fillers. We have also studied separation of complex gas mixtures such as CH₄/H₂ and N₂/CO₂ through these polymer-MOF MMMs performed using microsecond long all-atom nonequilibrium molecular dynamics (NEMD) simulations. Their simulation trajectories were analyzed in order to determine the effect of polymer rigidity and hydrophilicity, polymer chain length, and MOF geometry on their selectivity and permeability.

Active layer morphology plays a critical role in the performance of organic electronics. For most high performing material systems, the process of morphology formation involves simultaneous phase separation and crystallization of the acceptor layer materials. In order to fully understand the influence of processing conditions on the device performance, it becomes important to understand the formation of morphology under state-of-the-art device manufacturing techniques like solvents, co-solvents, additives and substrate induced phase separation. In this work, we first propose a meso-scale modeling framework to account for the effects of crystallization in solvent evaporation induced phase separation processes. This proposed framework has the capability to (a) be extensible to multiple crystallizing components, and (b) use experimentally measurable parameters like heat of crystallization and energy barrier of crystallization. We then demonstrate this multi-phase, multi-physics finite element framework for modeling crystallization and phase separation to a current high performing material system like PM6-Y6. PM6-Y6 system has been shown by several recent studies to have interesting crystallization-phase separation behavior, as well as a very high power conversion efficiency. Using this material system, we investigate the effect of evaporating conditions on the active layer morphology. Furthermore, we integrate this framework with virtual instruments to create an end-to-end platform for in-silico in-situ morphology characterization - quantification tool.

Microgels are highly compressible colloids that can swell and deswell by absorbing or expelling solvent from their interior. As such, suspensions of microgel particles can be concentrated far more than a suspension comprised of hard colloids. This leads to unusual mechanical properties of microgel suspensions including a drastic difference between the bulk and shear moduli. Microgels in compressed suspension can respond to compression in a variety of ways such as interpenetration, deformation, and shrinking. Previous experiments have offered insightful, but somewhat conflicting accounts of the behavior of individual microgels in compressed suspensions. We employ a mesoscale computational model to probe the mechanics and behavior of compressed microgel suspensions. We consider monodisperse microgels a viscous solvent and investigate the effects microgel architecture, the packing fractions, and solvent conditions on the suspension bulk modulus and the behavior of individual microgels. We use two methods to change the suspension packing fraction. In the first method, the suspension is volumetrically compressed, whereas in the second method the solvent quality is changed to induce microgel swelling and deswelling. We find that above random close packing microgels predominantly change the shape and mildly shrink. Microgel interpenetration remains relatively small and comparable to the mean distance between crosslinks. At higher packing fractions, microgels solely shrink. We find that the mechanics of suspension is defined by the bulk modulus of a single microgel. This result holds for microgels with different architectures, and insensitive to the method used to alter the suspension packing fraction even as the suspension undergoes colloidal gelation. Thus, the single-microgel bulk modulus established the correct mechanical measure for compressed microgel suspensions.

Molybdenum disulfide (MoS₂) is becoming increasingly popular as a substrate due to its more suitable electronic properties, higher signal-to-noise ratio, and increased biological compatibility as compared to graphene. Moreover, water is ubiquitous and most materials will inevitably come into contact with it. Consequently, understanding the structure of water at the MoS₂-water interface is imperative in determining the level of performance of MoS₂ devices in industrial and nanofluidic applications. Using experimental techniques, however, to understand interactions at the molecular level is too difficult using modern technologies. We have developed non-bonded FF parameters between an MoS₂ sheet and water molecules, which are represented by 12-6 Lennard-Jones (LJ) potential. The FF parameters were optimized to reproduce the macroscopic contact angle of water of ~69° on the MoS₂ sheet, which falls within the experimentally reported range of 65° to 75°. The MoS₂ sheet was represented by Stillinger-Weber (SW) potential. Three widely used water models, namely, SPC, SPC/E, and SPC/Fw, were used to model water. In order to validate the accuracy of the optimized FF parameters, the interaction energy of a single water molecule in its preferential orientation placed above a sheet of MoS₂ was calculated and compared to values reported in the existing literature. The structure of water at the MoS₂-water interface was also analyzed to validate the resulting parameters; specifically, the Z-density profile, average hydrogen bonds per molecule, and the angle made by the O-H vector with the normal to the MoS₂ sheet were calculated and studied.

Asphalt, used for roads and pavements, is an important building material that must endure a wide range of temperatures for many years. Asphalt originates from crude oil and contains millions of distinct organic molecules whose interactions contribute to the mechanical properties of the mixture. The study of this ubiquitous and complex material in essential in order to make its components more sustainable as well as improve durability to reduce the frequency of replacement. The effects of additives on asphalt have been extensively studied empirically, and several molecular models have arisen for the purpose of predicting interactions computationally. However, due to the chemical complexity of the system, there is still a large effort to improve current molecular models to understand interactions between different organic components within asphalt. In this study we firstly validate and refine the SARA (saturates, aromatics, resins, and asphaltenes) model [1] of asphalt in conjunction with experimental results. Secondly, we will use the improved model to describe the impact of additives on asphalt performance. Here we show a multistep process in which we utilize xTB, a semi-empirical tight binding program [2] to optimize molecular geometry, DFT for chemical property calculation, and molecular dynamics for large scale mechanical properties. We discuss how this approach of modeling complex molecular systems such as asphalt enables the rapid evaluation of several theoretical model compounds. We then use these results to propose potential additives for the improved sustainability and performance of asphalt.
References:
[1] Martín-Martínez, F.; J., Fini; E. H.; Buehler, M. J. RSC Advances 2015, 5(1), 753-759.
[2] Bannwarth, C.; Ehlert; S.; Grimme, S. Journal of chemical theory and computation 2019, 15, (3), 1652-1671.

Effects of framework deformation due to temperature and solvent (water) effects on the redox-active Co transition metal sites present in layered 2D conductive Co3(HTTP)2 metal-organic framework is studied in details. By combining classical molecular dynamics simulations with accurate and flexible ab initio parameterized force fields and ab initio periodic and cluster electronic structure calculations we show that including framework flexibility and solvent effects lead to higher reactivity for the oxidative dehydrogenation of propane to propene and hydrogen compared to the case that these effects are ignored in the conventional static electronic structure calculations.

Rigid polyurethane (PU) foams play crucial role in improving energy efficiency of our homes, offices, and appliances. The preparation of an insulating rigid PU panel is a complex physico-chemical process that includes several concurrent reactions (blowing and gelling), polymer molecular weight build-up, gelation and vitrification, bubble nucleation and growth, Ostwald ripening and coalescence; these phenomena take place as the reacting fluid is pushed into a mold and must occupy the space there without leaving any voids or other imperfections. Designing this process – from selecting the right formulation components (surfactants, catalysts, polyols, isocyanates, etc.) to optimizing the mold geometry – is a complex task requiring multiple and expensive trials. We describe a new multiscale modeling framework aiming to predict the foam flow into the mold and the final foam properties as a function of the formulation. The modeling tool has been validated in several experimental tests and is being utilized to help design new rigid PU foam formulations to better satisfy customer needs.

Molecular dynamics (MD) is a powerful approach to probe molecular-level mechanisms, thermodynamics, and kinetics, but is limited to millisecond time scales even on the world’s fastest supercomputers. We have developed molecular latent space simulators (LSS) to break this time scale barrier by learning highly efficient and accurate surrogate dynamical models of the long-time molecular dynamics. The LSS comprises three back-to-back deep learning networks with specialized architectures that are trained over modest MD training data to (i) learn an encoding of a molecular system into a slow latent space, (ii) construct a low-dimensional dynamic propagator in this space, and (iii) generatively decode molecular configurations to produce a synthetic molecular trajectory. The trained LSS is then deployed to generate novel ultra-long molecular trajectories at six orders of magnitude lower cost than MD. The low-cost of these synthetic trajectories enable resolution of rare thermodynamic states and kinetic transitions with arbitrarily low statistical uncertainties. In an application to Trp-cage mini-protein, we generate millisecond trajectories in just minutes of wall clock time, demonstrate excellent agreement with the structure, thermodynamics, and kinetics of the MD calculations, and resolve stabilities and rates with 10-fold higher accuracy.

Realistic simulations of polymers require a large number of atoms due to the anisotropic nature of their morphology. Although density functional theory (DFT) is considered highly accurate, it is too computationally expensive to handle large systems. Classical potentials used for molecular dynamics studies are trained on experimental or quantum mechanical equilibrium structures, and their accuracy away from equilibrium is not well known. This work is a comparative study of two widely used classical potentials—Optimized Potentials for Liquid Simulations (OPLS) and the reactive force-field (ReaxFF)—with respect to DFT. Their performance for polyethylene (PE), a simple model polymer, is benchmarked by comparing the classical energies, forces, and stresses against DFT.
The equilibrium crystal structure of PE is orthorhombic. In situations away from equilibrium, PE undergoes a phase transition to monoclinic. This work focuses on variations of these two crystal structures, along with some amorphous structures, to model realistic PE morphology. The energy resulting from the classical potentials is highly accurate when compared to DFT results. Force, an atomistic property, is not as easily captured, but stress, a bulk property, can also be accurately captured by classical potentials.
Additionally, a pressure-temperature phase diagram for PE is computed using OPLS and ReaxFF. ReaxFF performs better at high temperatures than OPLS when compared to experimental results. Although the classical potentials do not perfectly capture the DFT and experimental behavior, the results are close enough to justify their use when large system sizes are required. Consequently, these results will be used for shock simulations involving far from equilibrium structures.

Folding single polymer chains into nanoparticles mimics the natural process of folding peptides into proteins, whilst for single chain nanoparticles (SCNP), the combinatorial possibilities in polymeric components, precise control over intramolecular crosslinks, and tailored functionalization of the synthetic nanoparticle expand potentials beyond natural mechanism. Light-gated reversible photocycloaddition reactions allow bonds to form and break within the SCNP for controlled folding via gauging irradiation wavelengths.
Here we developed a holistic coarse-grained molecular dynamics model of photocycloaddition reactions within the confined environment of single polymer chains, which grants real-time monitoring capabilities in the degree of intramolecular crosslinking and morphology of the polymer coil. It is revealed that in the confined environment, cycloaddition and cycloreversion reactions lost orthogonal addressibiliy, when both shorter and longer wavelength irradiations induce SCNP folding. Separated in timescales between bond recombination and diffusion, their collective effect results in wavelength-gated photostationary states where degree of crosslinking and polymer chain compaction converge both from simulation and experiment.
The developed model of single-chain polymer folding serves as a blueprint in understanding the quantitative structure-property relationship and control of photostationary states within polymer networks.

Industrial applications of soft materials need to consider a whole range of additional factors in combination to scientific knowledge provided by modelling techniques alone. A proposed technical solution may fail if it turns out not to be economically viable. Therefore, it is vital to factor business logic into an organization’s decision making process when designing and producing new products.

The Formulations and Computational Engineering (FORCE) project, implemented within the EU Horizon 2020 program (https://www.the-force-project.eu), aims to bring materials modelling to the heart of business decision making via creation of Business Decision Support Systems (BDSS). By doing so, industries will be able to make more knowledgeable decisions based not only on existing legacy data, but also on new data generated by state of the art multi-scale and multi-physics materials modelling. Development of the FORCE BDSS and Workflow Manager has been closely aligned to activities of the European Materials Modelling Council (EMMC) and also builds on the SimPhoNy framework (https://www.simphony-project.eu) for multi-scale modelling in materials science.

The FORCE BDSS provides three key features:

- A library containing wrappers for state-of-the-art MCO algorithms
- A platform for designing and running MCO calculations on arbitrary systems
- A framework for integrating existing modelling and experimental pipelines together via a common universal data typing system

The interoperability of data between both modelling software and physical domains is placed at the core of our FORCE BDSS, which we demonstrate as the FORCE Workflow Manager GUI, available at https://github.com/force-h2020. This software provides an integrated framework that connects simulation models of varying complexity, experimental data sets and commercially relevant information together as a 'Workflow' pipeline construct. Consequently, users are able to select from a wide variety of parameters and key performance indicators (KPI) that can be used to identify the optimal trade off when making a business decision. For example, a set of KPIs may include material properties such as tensile strength or viscosity, but also component cost and availability. An integrated set of multi-criteria optimization (MCO) algorithms will then identify optimal outcomes to consider, typically by mapping out a Pareto frontier.

We report on an example use case targeting the industrial application of micelles that has been investigated using the software. By doing so, we demonstrate how the BDSS may be used to design simple pipelines that optimize chemical formulations by performing and processing molecular dynamics simulations. Our example MCO includes conflicting KPI goals that aim to yield the highest quality product under typical business constraints. Therefore we do not expect to be able to identify an optimal solution, but by exploring the Pareto frontier we attempt to locate the best set of trade offs that could be made for any given scenario.

Finally, we describe how the FORCE BDSS is not restricted as a modelling framework, but rather can be integrated into both data science and experimental laboratory pipelines. Future use cases aim to include Workflow constructs working with legacy datasets and containing a combination of simulation, machine learning and wet lab operations. By doing so, we open up the potential for machine learning and multi-scale modelling to be truly incorporated within an organization's R&D.

Molecular dynamics simulations use statistical mechanics at the atomistic scale to enable both the elucidation of fundamental mechanisms and the engineering of matter for desired tasks. The behavior of molecular systems at the microscale is typically simulated with differential equations parameterized by a Hamiltonian, or energy function. The Hamiltonian describes the state of the system and its interactions with the environment. In order to derive predictive microscopic models, one wishes to infer a molecular Hamiltonian that agrees with observed macroscopic quantities. Recent advances in automatic differentiation provide possibilities for the optimization and control of complex Hamiltonians with analytical gradients. From the perspective of engineering, one wishes to control the Hamiltonian to achieve desired simulation outcomes and structures, as in self-assembly and optical control, to then realize systems with the desired Hamiltonian in the lab. In both cases, the goal is to modify the Hamiltonian such that emergent properties of the simulated system match a given target. We demonstrate how this can be achieved using differentiable simulations where bulk target observables and simulation outcomes can be analytically differentiated with respect to Hamiltonians, opening up new routes for parameterizing Hamiltonians to infer macroscopic models and develop control protocols.

REACTER (www.reacter.org) is a heuristic protocol that enables the simulation of complex reactions using atomistic molecular dynamics (MD) with a fixed-valence force field. Incorporating reactions into classical MD with this approach allows modeling of reactive systems over greatly-increased time scales, enabling systems to be modeled with MD that would not otherwise be feasible. One or more competing multi-step reactions or series of reactions can be invoked simultaneously. Special treatment can be applied to neighboring atoms to relax high energy configurations while the simulation progresses. The original version of REACTER, which was implemented in the open-source LAMMPS simulation package as fix bond/react, was only available for serial simulations. This work describes the expansion of fix bond/react for use in parallel simulations, as well as the addition of various new options, including deletion of reaction by-products, reversible reactions, and custom reaction constraints. These new capabilities are demonstrated through large-scale simulations (200,000+ atoms) of the polymerization of polystyrene and nylon 6,6. The morphologies of both polymers are analyzed after reaching >99% extent of polymerization. Finally, the newly-added reversible reactions feature is demonstrated by rupturing these highly-entangled systems under uniaxial strain by defining a chain scission reaction.

In this talk, I will present our group’s effort to build a high-throughput infrastructure for the automated execution of materials science computations and analysis. This work is an extension of open-source Python packages that have been developed in Materials Project [1]: (1) pymatgen [2] for structure representation and input/output files generation and handling, (2) FireWorks [3] for managing workflows over computing resources, and (3) Custodian [4] for monitoring inevitable errors during the simulations and applying on-the-fly fixes. In our work, we benefit from these three libraries to build well-tested workflows for speeding up the prediction and screening of materials properties relevant to various applications with a focus on battery systems. At the backend, the infrastructure interfaces with the Gaussian [5] software which enables electronic structure calculations of chemical systems, and LAMMPS [6] open-source code for molecular dynamics (MD) simulations. In addition to its ability to handle the generation of input/data files and parsing of output files from the mentioned computational software, the infrastructure allows management of the collected data and storing it in MongoDB [7], a NoSQL database program using JSON-like documents with flexible schema. Examples of implemented Gaussian-based workflows include the calculation of electrostatic partial charges, NMR chemical shift, binding energy, ionization potential, electron affinity, and bond dissociation energy. Each derived property is saved in its own collection with auxiliary information like molecular metadata (smiles representation, chemical formula, …), which makes it easy to query and data-mine structure-property relationships. The user can tune the calculations by overriding default workflow parameters, for example, the functional and basis set, by-passing selected steps, or packing many jobs over multiple nodes for supercomputing resources. LAMMPS workflows allow the execution of MD simulations in different ensembles and analysis of the dumped trajectories for various dynamical and structural properties. The infrastructure enables coupling first-principles calculations with classical MD simulations to address chemical and physical phenomena occurring in a given system at different length and time scales. This approach aids in the development of databases required for training machine-learning models to accelerate the design and screening of materials with optimized properties for various materials science applications.

1. Jain, A., et al., Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. Apl Materials, 2013. 1(1): p. 011002.
2. Ong, S.P., et al., Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis. Computational Materials Science, 2013. 68: p. 314-319.
3. Jain, A., et al., FireWorks: a dynamic workflow system designed for high-throughput applications. Concurrency and Computation: Practice and Experience, 2015. 27(17): p. 5037-5059.
4. Custodian. <https://github.com/materialsproject/custodian>.
5. Frisch, M.J., et al., Gaussian 16 Rev. C.01. 2016: Wallingford, CT.
6. Plimpton, S., Fast parallel algorithms for short-range molecular dynamics. 1993, Sandia National Labs., Albuquerque, NM (United States).
7. MongoDB Inc., M., 2014.

Free energy and other related properties are not averages of functions of the phase space coordinates of the system, which means that they can not be calculated directly from the simulation trajectory. Under this category we find the chemical potential, which can be calculated with the Widom [4] particle insertion method. One drawback of this method is that it becomes very inefficient and often fails at higher densities. We propose a new method which utilizes the size and shape dependency of small systems to extract chemical potential differences from molecular dynamics simulation trajectories. The method is based on the difference in behavior of small and large systems, which can be summarized by the surface area to volume ratio. Since this ratio is much larger for small systems, the surface effects become much more significant, and the properties can no longer be directly compared to thermodynamic properties of macroscopic systems [1]. The Small System Method (SSM) [2], exploits the size dependency of these properties, and uses properties calculated for small systems to obtain the value of the same property for a macroscopic system. The small systems are created by placing small non-periodic subsystems at random positions inside a larger periodic reservoir. From statistical mechanics, the fluctuations in number of particles is coupled to different properties of the systems, like compressibility, heat capacity, partial molar volume and enthalpy. By combining SSM with the Overlapping Distribution Method originally developed by Bennett [3] we can also obtain values for chemical potential, pressure and free energy in the small systems. As with the previously mentioned properties, these can be extrapolated to the thermodynamic limit in order to obtain the macroscopic value of the properties. The method works well for a range of different densities, which makes it a promising alternative to the Widom [4] particle insertion method.


[1] Hill TL. Thermodynamics of Small Systems. J Chem Phys. 1962;36(12):3182–3197.
[2] Schnell SK, Vlugt TJH, Simon JM, Bedeaux D, Kjelstrup S. Thermodynamics of a small
system in a μT-reservoir. Chem Phys Lett. 2011;504(4):199 – 201.
[3] Bennett CH. Efficient estimation of free energy differences from Monte Carlo data. J Comput Phys. 1976;22(2):245-268
[4] WidomB. Some Topics in the Theory of Fluids. J Chem Phys. 1963;39(11):2808–2812.

There has been a recent surge in interest in soft and biological materials due to their potential applications in diverse areas including biomedicine, energy and nanoelectronics. The chemical diversity in the building blocks/sequences of the molecules constituting these materials yield an enormous molecular parameter space. A fundamental understanding of the impact of chemical block/sequence on the structure-activity relation of these soft and biological materials will accelerate their development and use, thereby accelerating the advancement of diverse disciplines.
Cyber-enabled computational methods, which integrate molecular simulations with analytical techniques, advanced computing and sampling tools, are used to study and develop new soft and biological materials with desired sequence/block-structure-activity properties in a more time- and cost-efficient way. The extended spatiotemporal scales connecting molecules to materials is addressed via the use of coarse-grained representations of the molecular species. The dynamics underlying the formation of the materials is resolved via the classical Molecular Dynamics simulation technique.

Topology and curvature changes in cellular membranes, especially in response to protein inclusions is important for a variety of cellular functions. Continuum elastic models such as the Helfrich model is able to describe several membrane elastic properties, but specific lipid-protein interactions (and feedback effects) are usually missing from these models. Here, we use coarse-grain molecular dynamics simulations to analyse the molecular-level mechanistic properties of a series of bilayers together with caveolin-1 as a protein inclusion . We show that in asymmetric, multi-component bilayers, differential stress between the two leaflets exists and may lead to non-additive curvature effects. The differential bilayer stress and the area mismatch cannot always be compensated directly in these systems. Binding of caveolin-1 can lead to a lipid clustering and de-mixing of multi-component bilayers. We show that the de-mixing is dependent both on specific protein-lipid interactions, as well as curvature-dependent lipid sorting. Our work is an important step in understanding mechanical properties of complex asymmetric bilayers and the role of protein inclusions.

References:
Krishna & Sengupta, 2019, Biophys. J.
Krishna et al., 2020, J. Phys. Chem. B

Peptide based biomaterials have garnered significant attention due to their ability to respond to external stimuli such as heat, light, and pH, thus allowing for controllable release of therapeutics for drug delivery applications and tunable mechanical and chemical properties for tissue engineering and in vitro cell culture applications. In this talk, I will present our recent work involving atomistic and coarse-grained molecular dynamics (MD) simulation studies of thermoresponsive polypeptides, specifically elastin-like peptides (ELP) and elastin-like peptide – collagen-like peptide (ELP-CLP) bioconjugates. ELPs are polypeptides that mimic the extracellular matrix protein, elastin, and are composed of repeat units of (VPGXG), where V, P, and G, are valine, proline, and glycine, respectively. The fourth residue, X, is termed the guest residue and can be any amino acid besides proline. Aqueous solutions of ELPs undergo a lower critical solution temperature (LCST)-like phase transition, wherein ELPs remain soluble below the transition temperature, T_t, and insoluble above T_t. We have focused our recent computational work on understanding the impact of substitution of F with Tyrosine (Y) and Tryptophan (W) guest residues and the sequence of these substitutions on the T_t of short ELPs in free state and when ELPs are conjugated to CLPs. Through a combination of atomistic and coarse-grained MD simulations we have described the molecular interactions that drive the observed shifts in the transition temperatures in our experimental collaborator's lab.

Eumelanin is an amorphous insoluble pigment present in both mammals and invertebrates. This natural material consists of amorphous aggregates of DHI (5,6-dihydroxyindole) and DHICA (5,6-dihydroxyindole-2-carboxylic acid) oligomers, interacting via π-π stacking and hydrogen bonding.¹ Synthetic melanins are usually obtained using either DHICA or DHI and present a lower degree of structural disorder. Due to their dual protonic and electronic conductivity^2,3 and intrinsic biocompatibility, melanins are being investigated for energy storage/production⁴ and as composite materials for biomedical applications.⁵
We present a systematic study of the electronic structure and conformational space of DHICA oligomers, characterized by a linear polymeric structure and strong antioxidant properties. Our computational insight not only fits well within existing experimental evidence, but can also predict the charge transport landscape in the bulk polymer.⁶ We then use molecular dynamics simulations to study the fiber-like structures formed by this material, and investigate its morphology dependent properties.
Finally, we describe our approach towards the simulation of realistic DHI melanin aggregates using a combination of experimental evidence, chemical kinetics and molecular dynamics.

1. d’Ischia, M.; Wakamatsu, K.; Cicoira, F.; Di Mauro, E.; Garcia-Borron, J. C.; Commo, S.; Galván, I.; Ghanem, G.; Kenzo, K.; Meredith, P.; et al. Pigment Cell Melanoma Res. 2015, 28 (5), 520–544.
2. Migliaccio, L.; Manini, P.; Altamura, D.; Giannini, C.; Tassini, P.; Maglione, M. G.; Minarini, C.; Pezzella, A. Front. Chem. 2019, 7 (March), 1–8.
3. Wünsche, J.; Deng, Y.; Kumar, P.; Di Mauro, E.; Josberger, E.; Sayago, J.; Pezzella, A.; Soavi, F.; Cicoira, F.; Rolandi, M.; et al. Chem. Mater. 2015, 27 (2), 436–442.
4. Kumar, P.; Di Mauro, E.; Zhang, S.; Pezzella, A.; Soavi, F.; Santato, C.; Cicoira, F. J. Mater. Chem. C 2016, 4 (40), 9516–9525.
5. Migliaccio, L.; Altamura, D.; Scattarella, F.; Giannini, C.; Manini, P.; Gesuele, F.; Maglione, M. G.; Tassini, P.; Pezzella, A. Adv. Electron. Mater. 2019, 5 (3), 1–8.
6. Matta, M.; Pezzella, A.; Troisi, A. J. Phys. Chem. Lett. 2020, 11, 3, 1045-1051.

Hierarchical multiscale modeling is used in this work to model collagen based biotissues, with direct application to the mechanics of the facet capsule ligament (FCL) of the spine. The biomaterial is modeled in the continuum sense using finite elements, whose constitutive description is provided by representative volume elements (RVEs) containing discrete representation of the underlying network. This technique offers a mechanism to predict the macroscopic properties of biological structures from the properties of the constituent fibers. A requirement of the hierarchical multiscale scheme is that RVEs at each material point must be much smaller than the corresponding continuum scale element. In fibrous materials, this can be problematic due to large size effects in three dimensions. Generalized boundary conditions are developed to mitigate the size effect and reduce the required size of the RVEs, therefore enabling larger multiscale models with increased macroscale geometry resolution. Predictions of the mechanical behavior of the FCL which account for the fiber-scale, microscale and ligament-scale collagen architecture are presented.

Self-affine morphology of random interfaces governs material functionalities across tribological, geological, (opto-)electrical and biological applications. However, the knowledge of how energy carriers or generally classical/quantum waves interact with structural irregularity is still incomplete. In this work, we study vibrational energy transport through random interfaces exhibiting different correlation functionals on the two-dimensional hexagonal lattice. We show that random interfaces are Cantor-composites populated on geometrical fractals, thus multifractals, and calculate their quantized conductance using atomistic approaches. The interfacial conductance increases when disorder enhances, and we obtain a universal scaling law. This scaling law contains self-similarity for mass perturbation, and exponential scaling of structural irregularity quantified by Hausdorf dimension. We resolve the modal characteristics by atomic Green’s function, wavepacket simulation, and molecular dynamics, and ascribe the disorder-induced enhancement of conductance to the increased harmonic transmission of transverse and flexural modes, augmented by anharmonic interactions near interfaces. The multifractal nature and Cantor-composite picture may be extendable to charge and photon transport across random interfaces.

Materials informatics, and the associated field of nanoinformatics, offer a plethora of new approaches to solving existing challenges in (nano)materials design. These enabling technologies leverage over 50 years of innovation in theoretical computer science on machine learning algorithms, and an even longer history of research in statistics to underpin the essential preliminary data science. Provided sufficient appropriate data is available to describe the material or system, new insights can be gained that would be otherwise obscured using conventional experimental or computational methods. Hidden structure/property relationships we can use to inform further research and materials development. Extracting useful insights (such as structure/property relationships) from data analytics and machine learning is however more complicated than simply gathering data and training models, since not all methods are interpretable, and interpretability is essential for decision making. Correlation is also different to causation, and so the importance relationships identified using many machine learning methods are not always actionable. In this presentation we will discuss the importance of accuracy, generalizability and interpretability, and demonstrate the advantages of combining a series of different machine learning methods to uncover useful relationships between the properties, structure and processing conditions of graphene oxide nanoflakes.

The ever-increasing power of modern supercomputers, along with the availability of highly scalable atomistic simulation codes, has begun to revolutionize predictive modeling of materials. Molecular dynamics (MD), in particular, has led to breakthrough advances in diverse fields, including tribology, energy storage, catalysis, sensing. Furthermore, recent integration of MD simulations with X-ray characterization has demonstrated promise in real-time 3-D atomistic characterization of materials. The popularity of MD is driven by its applicability at disparate length/time-scales, ranging from ab initio MD (hundreds of atoms and tens of picoseconds) to all-atom classical MD (millions of atoms and tens of nanoseconds), and coarse-grained (CG) models (microns and tens of micro-seconds). Nevertheless, a substantial gap persists between AIMD, which is highly accurate but restricted to extremely small sizes, and those based on classical force fields (atomistic and CG) with limited accuracy but access to larger length/time scales. The accuracy and predictive power of classical MD is dictated by the empirical force fields, and their capability to capture the relevant physics. In this talk, we will present some of our recent work on the use of machine learning (ML) to seamlessly bridge the electronic, atomistic and mesoscopic scales for materials modeling. Our automated ML framework aims to bridge the significant gulf that exists between the handful of research groups that develop new interatomic potential models (often requiring several years of effort) and the increasingly large user community from academia and industry that applies these models. Our ML approach has showed marked success in developing force fields for a wide range of materials from metals, oxides, nitrides, hetero-interfaces to two-dimensional (2-D) materials and even water (arguably the most difficult system to capture from a molecular perspective). This talk will also briefly discuss our ongoing efforts to integrate such cheap yet accurate atomistic models with (a) AI techniques to perform inverse design and construct metastable phase diagrams of materials (b) Deep learning to improve spatiotemporal resolutions of ultrafast X-ray imaging.

Developments in machine learning have opened the door to a new generation of methods for studying phase transitions, due to their ability to efficiently and autonomously identify complex patterns in many-body systems. Applications of machine learning techniques in statistical physics vary from locating phase transitions in spin systems to pinpointing weak spots in colloidal glasses. The rapid emergence of machine learning applications in both statistical mechanics and materials science demonstrates that these techniques are destined to become an important tool for soft matter physics.

One problem this tool seems ideally suited to solve, both experimentally and in simulations, is the identification of different (crystalline) phases – even on the level of a single particle. For example, when we use either experiments or simulations to look at crystallization at the microscopic level, it is always a challenge to determine which particles are part of the fluid, and which ones have already crystallized. In this talk I will describe how machine learning techniques can automatically figure this out for us, even if we do not know in advance what crystal structure to look for. We will also examine how such methods might reveal new structural insights in “purely” disordered systems like glasses.

Inverse problems in image processing, phase imaging, and computer vision often share the same structure of mapping input image(s) to output image(s) but are usually solved by different application-specific algorithms. Deep convolutional neural networks have shown great potential for highly variable tasks across many image-based domains, but sometimes can be challenging to train due to their internal non-linearity. We propose a novel, fast-converging neural network architecture capable of solving generic image(s)-to-image(s) inverse problems relevant to a diverse set of domains. We show this approach is useful in recovering wavefronts from direct intensity measurements, imaging objects from diffusely reflected images, and denoising scanning transmission electron microscopy images, just by using different training datasets. These successful applications demonstrate the proposed network to be an ideal candidate solving general inverse problems falling into the category of image(s)-to-image(s) translation.

Adhesion is a ubiquitous phenomenon that exists between every pair of surfaces. As an interfacial force, adhesion becomes dominant at nanoscale compared with body forces. Nanoscale adhesion is critical for applications such as the design of nanoscale devices, MEMS, and nanofabrication. Local roughness at the nano- and atomic-scales significantly influence the adhesion and can reduce the adhesion by more than one order of magnitude than that predicted by classic ideallized models. Statistical parameters, such as root mean square (rms) roughness, are proposed to develop new adhesion models and characterize nanoscale adhesion properties. Although the roughness is simplified into a statistical parameter in this way, it is inevitable to lose a lot of useful details of topological data. Here, we use machine learning to characterize interfacial adhesion properties based on nanoscale topology-adhesion correlated data. We perform high frequency quantitative nanoscale mechanical mapping (QNM) by AFM to simultaneously acquire topology at sub-nm resolution and adhesion at nanoscale at the interface between an AFM tip and a nominally flat surface. The adhesion varies over 20 folds (7.57 nN to 166.4 nN) due to nanoscale roughness. Local adhesion data is used as target and topological data at this and nearby locations (in-plane radius of 0.5 to 8 nm) is used as features. We use several machine learning methods to analyze the data, among which support vector regression (SVR) with radial basis function (RBF) kernel has the optimal performance. The prediction of the testing set (40% training and 60% testing data set) reaches r² more than 0.97. Based on the results of machine learning, we develop an interfacial mechanics model that establishes the relationship between detailed local topology and adhesion. This model can be generalized for different materials and interfaces to understand nanoscale adhesion and the effect of roughness.

The dielectric constant (ε) is a critical parameter utilized in the design of polymeric dielectrics for energy storage capacitors, microelectronic devices, and high-voltage insulations. However, agile discovery of polymer dielectrics with desirable ε remains a challenge, especially for high-energy, high-temperature applications. To aid accelerated polymer dielectrics discovery, we have developed a machine-learning (ML)-based model to instantly and accurately predict the frequency-dependent ε of polymers with the frequency range spanning 15 orders of magnitude. Our model is trained using a dataset of 1210 experimentally measured ε values at different frequencies, an advanced polymer fingerprinting scheme and the Gaussian process regression algorithm. The developed ML model is utilized to predict the ε of synthesizable 11,000 candidate polymers across the frequency range 60–10¹⁵ Hz, with the correct inverse ε vs. frequency trend recovered throughout. Furthermore, using ε and another previously studied key design property (glass transition temperature, Tg) as screening criteria, we propose five representative polymers with desired ε and Tg for capacitors and microelectronic applications. This work demonstrates the use of surrogate ML models to successfully and rapidly discover polymers satisfying single or multiple property requirements for specific applications.

Phase segregation of binary mixture has many applications including their popular application to investigate the evolution bulk heterojunction morphology (BHJ) from solvent evaporation of ternary mixture consisting of donor, acceptor and solvent molecules. Typically, macroscopic time evolution of phase separation is modeled computationally using system of coupled partial differential equations (PDE). Spontaneous separation of initially dispersed binary mixture is predominantly approximated by Cahn-Hillard equations. However, the numerical simulations using physics-driven PDE equations are iterated through a computationally expensive scheme until the final steady-state decomposition is achieved. On the other hand, deep learning techniques enabled great progress in the area of computer vision applied to microstructure evolution. We employ convolutional neural network (CNN) to predict phase segregation by sampling final decomposition of binary mixture purely based on observations from coarse-grained molecular dynamics simulations. Our deep learning approach can learn and predict non-linear physical phenomenon such as phase separation without any knowledge of the underlying atomic interactions between the constituent materials. The ability to infer phase segregation phenomena is potentially appealing for the rational microstructure design given several solution processing conditions, the initially dispersed microstructure consisting of the constituent materials and their interaction parameters.

The presence of defects in two-dimensional (2D) materials, such as graphene, can deteriorate the properties, performances, and functionalities of graphene-based devices. The prior knowledge of defect information in graphene such as the defect type, size, and density allows us to make predictions of properties of graphene. Nevertheless, the detection of unknown defects with an atomic-resolution probe such as transmission electron microscope is technically demanding and involves complex sample preparations. An alternative perspective arises when the thermal vibration of graphene sheets is examined. Because the absence or the substitution of an atom alters the boundary condition of adjacent oscillators, defective regions exhibit different vibrational properties compared to the pristine counterpart. However, the relation between defects and thermal vibration properties cannot be modeled explicitly. Here, we develop a machine learning-based approach to predict the unknown defect locations by thermal vibration topographies of graphene sheets, which overcomes the implicit relation between defect contents and vibration properties. Based on a kernel ridge regression model, two prediction strategies are developed, an atom-based method constructing data by atom indices, and a domain-based method constructing data by domain discretization. Trained by hundreds of thousands of vibrational energy distributions computed by molecular dynamics simulation, both methods can detect unknown graphene defects with a high accuracy. While the atom-based method is capable to predict a single-atom vacancy, the domain-based method can predict an unknown number of defects with an arbitrary distribution. The proposed data-driven strategy may also shed light on the defect detection in 2D materials of a broad variety.

Polymers are an important and diverse class of materials that span a huge chemical and property space. Machine learning methods have been recently successfully deployed to explore this space, revealing previously unidentified and novel polymers. The training of machine learning models requires a numerical representation of polymers, commonly termed fingerprints, as inputs which are “mapped” to the polymer properties as outputs. Single-task machine learning models learn the mapping between fingerprints and individual properties, one at a time. In contrast, multi-task algorithms learn the mapping of the chemical structure to multiple properties simultaneously. The effectiveness of this multi-task approach is because cross-property correlations, when present, are exploited implicitly. In this work, we benchmark two neural network based multi-task models against two single-task models for 27 different polymer properties. All models are trained on a polymer database containing ~8,000 polymers encompassing ~16,000 property points. The multi-task models demonstrate superiority over the single-task models, specifically for correlated polymer properties. An interpretability analysis of the multi-task model has also been performed that allow for the extraction of structure-property design rules.

Recent advances in machine learning have inspired the development of novel coarse grained molecular dynamics techniques wherein large length and time scale systems can be made more computationally tractable. Such regimes are essential for studying the dynamics soft materials, protein folding, and other complex processes that, in most cases, cannot be treated with current fully atomistic methods. In these schemes, where degrees of freedom in the all-atom picture are integrated out, one of the primary challenges lies in the construction of a free energy surface.

While the learned effective potential can be constructed on a variety of different physical bases, we consider those potentials that preserve the statistical mechanical properties of the atomistic system [1]. It is well known that the coarse graining process can induce complexity in the coarse grained potential that is not present in the all-atom case. In this work, we demonstrate the applicability of Gaussian processes for determining the optimal complexity of the free energy surface. This knowledge is crucial in designing models that are both computationally efficient and capture the essential physics of the problem.

In addition, we show the utility of active learning in the context of coarse graining. Recent efforts have shown that force field construction for ab initio molecular dynamics with Gaussian process regression [2] is amenable to active learning frameworks, and we demonstrate how this idea is transferable to coarse grained potentials. In this setting, we address the issue of “fine graining,” wherein an atomic representation of a given latent space despite the dramatic loss of information may be desirable.

[1] W. G. Noid, Jhih-Wei Chu, Gary S. Ayton, Vinod Krishna, Sergei Izvekov, Gregory A. Voth, Avisek Das and Hans C. Andersen. “The multiscale coarse-graining method. I. A rigorous bridge between atomistic and coarse-grained models,” J. Chem. Phys. 128, 244114 (2008).
[2] Jonathan Vandermause, Steven B Torrisi, Simon Batzner, Alexie M Kolpak, and Boris Kozinsky. Submitted. “On-the-Fly Bayesian Active Learning of Interpretable Force-Fields for Atomistic Rare Events,” npj Computational Materials 6 (2020).

Polynorbornene (PNB) is an important amorphous polymer system, which has potential applications as a high energy density polymer due to its high breakdown strength with low dielectric loss and high thermal stability. Moreover, electrical properties of PNB can be significantly enhanced by incorporation of defects or synthesis with controlled crystallinity by hydrogenation reaction. However, this process is challenging, since it involves experimental synthesis and characterization of combinatorially large number of polymer systems to identify potential candidates. Here, we propose a deep learning-based graph neural network (GNN) model with attention mechanism that can identify polymer systems capable of exhibiting increased energy and power density. The GNN model is trained to predict dielectric constant for a polymer, where the training data for the high-frequency dielectric constant of PNB polymers are computed via ab-initio molecular dynamics simulation. The attention mechanism in our GNN model provides key insight about the functional groups and their spatial location inside the polymer chain that contributes the most in its dielectric properties. Our model can significantly aid experimental synthesis of potentially new dielectric polymer materials, which is otherwise difficult using simplistic statistical procedures.

*This work was supported by the Office of Naval Research through a Multi-University Research Initiative (MURI) under grant number (N00014-17-1-2656).

Deep Learning has great promise for accelerating Molecular-Dynamics simulations by learning complex representations of atomistic data. We present Tensor-Field Molecular Dynamics, an SE(3)-equivariant Graph-Neural Network that leverages rotationally equivariant graph convolutions to operate on geometric tensors (rather than only scalar values) to learn energies and forces of atomic systems. We demonstrate state-of-the art performance on a diverse series of systems, including molecules, complex surface reactions, solvated polymers, extended systems, and water. In addition, we show that our proposed methods generalizes well from very small training data sets, which challenges the common belief that deep neural networks require massive training sets for accurate predictions. We then analyze a series of resulting Molecular Dynamics Simulations and compare thermodynamic and kinetic properties to the original ab-initio simulations. Finally, quantifying uncertainty of model prediction is critical for reliable long dynamics simulations. We outline a scalable method to obtain predictive uncertainty estimates of energies and forces from Tensor-Field models and present results on several different systems.

Many organic electronics applications such as organic solar cells or thermoelectric generators rely on PEDOT:PSS as a conductive polymer that is printable and transparent. It was found that doping PEDOT:PSS with sorbitol enhances the conductivity through morphological changes. However, the microscopic mechanism is not well understood. We combine multiscale modeling techniques with machine learning to investigate changes in morphological and electronic properties of PEDOT:PSS when doped with sorbitol. We find that sorbitol improves the alignment of PEDOT oligomers, leading to a reduction of energy disorder and an increase in electronic couplings between PEDOT chains. The high accuracy (r² > 0.9) and speedup of energy level predictions of neural networks compared to density functional theory enables us to analyze HOMO energies of PEDOT oligomers as a function of time which was computationally infeasible without machine learning. We find a surprisingly low degree of static energy disorder compared to other organic semiconductors. This finding might help to better understand the microscopic origin of the high charge carrier mobility of PEDOT:PSS in general and potentially help to design new conductive polymers.[1]
We will furthermore show how the same machine learning approach can be applied to other organic semiconductors to evaluate static and dynamic energy disorder on an level of detail that was previously impossible due to the large computational cost of quantum mechanical calculations.

[1] The influence of sorbitol doping on aggregation and electronic properties of PEDOT:PSS: a theoretical study, Pascal Friederich, Salvador Leon, Jose Dario Perea Ospina, Loic Roch and Alan Aspuru-Guzik, Machine Learning: Science and Technology, 2020

the relationship, if any, between the structure and creep dynamics of gels remains elusive. Here, based on accelerated molecular dynamics simulations and classification-based machine learning, we show that the propensity of a gel to exhibit long-time creep is encoded in its static, unloaded structure. We extract a local, non-intuitive structural descriptor (“softness”) that is strongly correlated with the dynamics of the gel particles—wherein the macroscopic creep rate exhibits an exponential dependence on the average softness. We find that creep results in a decrease in softness in the gel structure, which, in turn, explains the gradual decay of the creep rate over time.

Energy storage systems such as lithium-ion batteries (LIB) are ubiquitous in modern society with particular applications from consumer electronics to electric vehicles. One of the most fundamental challenges of studying and modeling the energy storage systems stems from the high dimensionality of the system, known as the curse of dimensionality (a great number of variables or degrees of freedom). For a commercial liquid-electrolyte battery cell, a comprehensive theoretical characterization of the LIB system should incorporate all the governing equations of the mechanics of the solid phases (electrodes, separator, and current collectors) and the liquid phase (electrolyte), the electrochemistry that involves mass transfer process of lithium ions, kinetics (rate of the reactions), and thermodynamics of materials, as well as the heat transfer processes over the electrodes and the whole battery cell. The complexity of the system increases exponentially considering that each of the governing equations has its own initial and boundary conditions that depend on the external loads and charge-discharge battery cycling behavior. Analytically solving such a system is almost impossible, and the real-world applicability of the existing so-called multi-physics numerical models based on the finite element method is usually hurt by the harsh conditions for convergence. Here, we use physics-guided machine learning algorithms to solve the coupled partial differential equations (PDE) of energy storage systems, and in this study, we demonstrate the applicability of the developed tool by using it to solve the poro-plasticity behavior of the battery, which involves Drucker-Prager/cap plasticity of the solid phase, Darcy’s law of the liquid phase, as well as the continuity equations. A neural network taking the spatial coordinates as the input and the displacement and stress of the solid phase and the velocity and pressure of the liquid phase as the output is constructed to approximate the exact solution. The prediction of the overall mechanical response is compared with the experimental data as validation. At last, a simplified battery electrochemistry model based on the multiphase porous electrode theory (MPET) is incorporated into the poro-plasticity framework, and the overall model is implemented by designing neural networks to reflect the multiple physics.

Organic thin films made of conjugated polymers and/or small molecules have captured the interest of the organic electronics industry due to their wide range of alterable properties (e.g., the color of light emission and solubility in organic solvents). High-performing devices have recently been designed with a multicomponent material system (as opposed to initial studies that focused on binary blends). This reaffirmed the need for a holistic approach to materials design beyond binary blends used to understand the mixing behavior and morphology formation during the fabrication of multicomponent organic thin films.

In this talk, we present a thermodynamics-based framework to identify the suitable solvents for a given electron-donor and accepting material with good performance as organic solar cells (OSC). We present high throughput exploration and construction of a material-phase diagram map that captures thermodynamics characteristics and facilitates understanding of the multicomponent solution's mixing behavior. Our analysis pipeline consists of (i) a physics based model to determine phase diagrams of multicomponent mixtures using the convex hull method; and (ii) high throughput analysis of a large set of phase diagrams using dimensionality reduction coupled with the clustering. To illustrate our pipeline, we derive the design rules for a few material systems typically used in OSC and compare it to the solubility sphere design rules.

Block copolymer (BCP) self-assembly, and in particular directed self-assembly (DSA), has provided a powerful route to fabricate complex thin film structures at small length scales. Despite DSAs success in fabricating various 2D patterns, fabrication of complex 3D nanostructures remains a challenge. Here, we will introduce a novel method, which we call Self-Directed Self-Assembly, and demonstrate how it can be leveraged to assemble 3D designer nanostructures using block copolymers. In particular we will demonstrate that uniform and non-uniform multilayer nanostructures can be obtained through stacking two or more "orthogonal" block copolymers. Finally, by introducing a patterned substrate to the first layer and performing multilayer stacking, we will demonstrate how to propagate information upwards in the film. Different complex multilayer structures will be shown, and at the end we will present an outlook on exciting new computational and artificial intelligence avenues of research in this field.

DNA-mediated assembly of colloidal particles can be utilized to produce a variety of structures which may have desirable phononic, photonic, or electronic transport properties. Recent developments in linker-mediated assembly processes allow for interactions to be coordinated between many different types of colloidal particles more easily and with fewer unique sequences than direct hybridization. However, the dynamics of programmed self-assembly become increasingly more complex when coordinating interactions between three or more distinct interacting elements. In such cases particle pairs with similar binding energies are allowed to interact unpredictably, and enthalpically degenerate binding sites will be noticeably more present while numerous secondary phases may also result from the self-assembly process. Therefore, it is necessary to develop procedures for predicting feasible superstructure geometries for these systems before they can be implemented in material design.

Here we investigate the formation of multi-component ordered structures through self-assembly of spherically symmetrical colloidal particles of different sizes with a variety of interaction matrices. We utilize Molecular Dynamics (MD) and Brownian Dynamics (BD) simulations to study the growth behavior of systems with different types of interacting elements and predict the formation and stability of target structures. Our results provide direct guidelines for designing multifarious colloidal structures with three or more types of building blocks. We also study phononic spectrum of various ternary structures in order to identify the influence of key structural parameters on phonon bandgap frequencies and ranges. Overall, this study provides new directions for future experimental work to target formation of multi-component colloidal superstructures beyond the well-established binary symmetries studied in the past.

Dissipative particle dynamics (DPD) has demonstrated excellent capabilities in describing the assembly and dynamics of soft-matter systems in the past. Its intuitive representation and ease of use has been important to many areas, and in particular to self-assembly of block copolymers (BCPs). However, DPD in its original implementation to block copolymers is unable to capture accurately the different entropies for block copolymers at interfaces, which leads to erroneous qualitative results for BCP thin films. We have recently re-parameterized the DPD force field to yield accurate results in a multitude of thin film conditions. Here we present some of the unanticipated consequences of such new parametrization that appears to predict much better experimental results for phase transitions in BCPs than the traditional field theoretic methods. In particular we will introduce results from different densities to show how the order-disorder transition converges to a universal value similar to the experimental value, how the order-disorder transition is affected by polymer length, and finally how one can combine directly evolutionary algorithms to predict grapho and chemo epitaxial templates for desired thin film morphologies. Our results indicate that for typical copolymers with molecular weights below ~10^5 g/mol, the order disorder transition occurs at a universal value of χN. Furthermore, our results clarify the role of entropic elasticity in the order-disorder transition. Simple scaling arguments demonstrate that SCFT and its corrections will fail to capture the correct order disorder temperature in the short to medium length copolymers. Our work should be useful to understand how experimentally relevant BCPs self assemble in different conditions as well as introduce the audience to new advances in the DPD technique that allow it to rival any field theoretic method, especially considering that DPD is a much more intuitive platform which is easier to implement and that it has a smaller computational footprint.

The formation of porous structures such as diamond by bottom-up assembly remains one of the significant open challenges in colloidal systems. The interest in these materials with functional architectures stems from their broad potential applicability in diverse fields such as catalysis, separations, and photonics. The recently proposed approaches based on “DNA-functionalized particles” (DFPs) show promising results in the synthesis of low-coordination open structures such as diamond lattices in nanoparticle assemblies [1,2]. However, the underlying design approach in most of these cases is usually too complex to be scalable. In this work, for the first time, we demonstrate using muti-scale molecular simulations that low-coordination lattices (e.g., honeycomb in two-dimensional (2D) and diamond in three-dimensional (3D) systems) can be formed using simple isotropic functionalization of nanoparticles with DNA. The strength of our approach lies in its simplicity and potential ease with which the proposed design scheme can be implemented in the laboratory experiments.

[1] Wang, Yifan, et al. "Colloidal crystals with diamond symmetry at optical lengthscales." Nature communications 8.1 (2017): 1-8.
[2] Liu, Wenyan, et al. "Diamond family of nanoparticle superlattices." Science 351.6273 (2016): 582-586.

Using solvent-free solid polymer electrolytes in batteries would reduce concerns about flammability and potentially decrease weight and cost, especially if their ion conductivity can be improved. Strategies to improve ion conduction in ion containing polymer materials include using bulky anions or using a high dielectric constant polymer. These both effectively decrease the strength of Coulomb interactions between ions at contact and can reduce ion agglomeration, which can improve conductivity. However, strong ion-polymer interactions can slow ion motion, and size asymmetry between ions may increase preferential solvation of cations versus the larger anions, lowering the transference number t₊ (fraction of conductivity contributed by the cation). We study these effects using coarse-grained molecular dynamics (MD) simulations which include a 1/r⁴ potential to capture size-dependent ion-monomer and ion-ion solvation effects. This is the same form as the interaction between an ion and an induced dipole. For block copolymer systems, we have also been guided by theoretical calculations to show the effects of macroscopic phase behavior. From MD, we calculate conductivity from ion mobilities in an external electric field, which improves accuracy versus the typical use of fluctuation dissipation relationships. We assess the impacts of ion size and polymer dielectric strength, including polymer dielectric strength contrast for block copolymer systems, on ion correlations, diffusion, and conductivity.

Proteins are known to undergo a dynamical transition at low temperature, typically between 220 and 240 K. This transition consists in a steep enhancement of the protein atomic mobility which is related to the onset of anharmonic motions. At the same temperature proteins start to be biologically active and thus a full understanding of the molecular origin of this phenomenon is of particular relevance. While the protein dynamical transition always occurs in aqueous environment, with the protein confinement preventing the onset of ice crystallization, the role of water is still a controversial issue.
Here we focus on the low temperature behavior of a non-biological macromolecule, i.e. poly(N-isopropylacrylamide), commonly known as PNIPAM. This polymer is mostly exploited for its thermo-responsive nature, but it is also of interest because of its affinity with proteins in terms of energy landscape and amphiphilic chemical composition. Depending on the specific synthesis protocol, different PNIPAM based molecular architecture can be originated, such as linear chains or microgels, i.e. polymer network particles with a colloidal size. In this talk I will discuss two studies on the low temperature behavior of this synthetic soft material where atomistic molecular dynamics simulations are combined to elastic incoherent neutron scattering experiments.
In the first part I will provide a description of the evidences of the occurrence of a low temperature dynamical transition in PNIPAM microgels, akin to that observed in proteins [1,2]. This study is based on a nanoscale model of a microgel network in water, which quantitatively reproduces neutron scattering experiments. By correlating the information extracted from the analysis of the polymer relaxations times, water self-diffusion coefficients and hydrogen bonding interactions I will show that water-polymer coupling plays a driving role in the phenomenon.
In the second part, I will report the observation of a low temperature dynamical transition also in PNIPAM linear chains, which suggest a wide generality of the phenomenon, independently on the macromolecular architecture [3]. I will also show that simulations allow to identify the specific motions underlying the onset of the transition.

References
[1] Zanatta M. et al. “Evidence of a low-temperature dynamical transition in concentrated microgels” Science advances 2018, 4, eaat5895.
[2] Tavagnacco L. et al. “Water-polymer coupling induces a dynamical transition in microgels” The journal of physical chemistry letters, 2019, 10, 870-876.
[3] Tavagnacco L. et al. “The dynamical transition of hydrated polymer chains” in preparation.

Rigid polyurethane (PU) foams play crucial role in improving energy efficiency of our homes, offices, and appliances. The preparation of an insulating rigid PU panel is a complex physico-chemical process that includes several concurrent reactions (blowing and gelling), polymer molecular weight build-up, gelation and vitrification, bubble nucleation and growth, Ostwald ripening and coalescence; these phenomena take place as the reacting fluid is pushed into a mold and must occupy the space there without leaving any voids or other imperfections. Designing this process – from selecting the right formulation components (surfactants, catalysts, polyols, isocyanates, etc.) to optimizing the mold geometry – is a complex task requiring multiple and expensive trials. We describe a new multiscale modeling framework aiming to predict the foam flow into the mold and the final foam properties as a function of the formulation. The modeling tool has been validated in several experimental tests and is being utilized to help design new rigid PU foam formulations to better satisfy customer needs.

Molecular dynamics (MD) is a powerful approach to probe molecular-level mechanisms, thermodynamics, and kinetics, but is limited to millisecond time scales even on the world’s fastest supercomputers. We have developed molecular latent space simulators (LSS) to break this time scale barrier by learning highly efficient and accurate surrogate dynamical models of the long-time molecular dynamics. The LSS comprises three back-to-back deep learning networks with specialized architectures that are trained over modest MD training data to (i) learn an encoding of a molecular system into a slow latent space, (ii) construct a low-dimensional dynamic propagator in this space, and (iii) generatively decode molecular configurations to produce a synthetic molecular trajectory. The trained LSS is then deployed to generate novel ultra-long molecular trajectories at six orders of magnitude lower cost than MD. The low-cost of these synthetic trajectories enable resolution of rare thermodynamic states and kinetic transitions with arbitrarily low statistical uncertainties. In an application to Trp-cage mini-protein, we generate millisecond trajectories in just minutes of wall clock time, demonstrate excellent agreement with the structure, thermodynamics, and kinetics of the MD calculations, and resolve stabilities and rates with 10-fold higher accuracy.

There has been a recent surge in interest in soft and biological materials due to their potential applications in diverse areas including biomedicine, energy and nanoelectronics. The chemical diversity in the building blocks/sequences of the molecules constituting these materials yield an enormous molecular parameter space. A fundamental understanding of the impact of chemical block/sequence on the structure-activity relation of these soft and biological materials will accelerate their development and use, thereby accelerating the advancement of diverse disciplines.
Cyber-enabled computational methods, which integrate molecular simulations with analytical techniques, advanced computing and sampling tools, are used to study and develop new soft and biological materials with desired sequence/block-structure-activity properties in a more time- and cost-efficient way. The extended spatiotemporal scales connecting molecules to materials is addressed via the use of coarse-grained representations of the molecular species. The dynamics underlying the formation of the materials is resolved via the classical Molecular Dynamics simulation technique.

Topology and curvature changes in cellular membranes, especially in response to protein inclusions is important for a variety of cellular functions. Continuum elastic models such as the Helfrich model is able to describe several membrane elastic properties, but specific lipid-protein interactions (and feedback effects) are usually missing from these models. Here, we use coarse-grain molecular dynamics simulations to analyse the molecular-level mechanistic properties of a series of bilayers together with caveolin-1 as a protein inclusion . We show that in asymmetric, multi-component bilayers, differential stress between the two leaflets exists and may lead to non-additive curvature effects. The differential bilayer stress and the area mismatch cannot always be compensated directly in these systems. Binding of caveolin-1 can lead to a lipid clustering and de-mixing of multi-component bilayers. We show that the de-mixing is dependent both on specific protein-lipid interactions, as well as curvature-dependent lipid sorting. Our work is an important step in understanding mechanical properties of complex asymmetric bilayers and the role of protein inclusions.

References:
Krishna & Sengupta, 2019, Biophys. J.
Krishna et al., 2020, J. Phys. Chem. B

Peptide based biomaterials have garnered significant attention due to their ability to respond to external stimuli such as heat, light, and pH, thus allowing for controllable release of therapeutics for drug delivery applications and tunable mechanical and chemical properties for tissue engineering and in vitro cell culture applications. In this talk, I will present our recent work involving atomistic and coarse-grained molecular dynamics (MD) simulation studies of thermoresponsive polypeptides, specifically elastin-like peptides (ELP) and elastin-like peptide – collagen-like peptide (ELP-CLP) bioconjugates. ELPs are polypeptides that mimic the extracellular matrix protein, elastin, and are composed of repeat units of (VPGXG), where V, P, and G, are valine, proline, and glycine, respectively. The fourth residue, X, is termed the guest residue and can be any amino acid besides proline. Aqueous solutions of ELPs undergo a lower critical solution temperature (LCST)-like phase transition, wherein ELPs remain soluble below the transition temperature, T_t, and insoluble above T_t. We have focused our recent computational work on understanding the impact of substitution of F with Tyrosine (Y) and Tryptophan (W) guest residues and the sequence of these substitutions on the T_t of short ELPs in free state and when ELPs are conjugated to CLPs. Through a combination of atomistic and coarse-grained MD simulations we have described the molecular interactions that drive the observed shifts in the transition temperatures in our experimental collaborator's lab.

Materials informatics, and the associated field of nanoinformatics, offer a plethora of new approaches to solving existing challenges in (nano)materials design. These enabling technologies leverage over 50 years of innovation in theoretical computer science on machine learning algorithms, and an even longer history of research in statistics to underpin the essential preliminary data science. Provided sufficient appropriate data is available to describe the material or system, new insights can be gained that would be otherwise obscured using conventional experimental or computational methods. Hidden structure/property relationships we can use to inform further research and materials development. Extracting useful insights (such as structure/property relationships) from data analytics and machine learning is however more complicated than simply gathering data and training models, since not all methods are interpretable, and interpretability is essential for decision making. Correlation is also different to causation, and so the importance relationships identified using many machine learning methods are not always actionable. In this presentation we will discuss the importance of accuracy, generalizability and interpretability, and demonstrate the advantages of combining a series of different machine learning methods to uncover useful relationships between the properties, structure and processing conditions of graphene oxide nanoflakes.

The ever-increasing power of modern supercomputers, along with the availability of highly scalable atomistic simulation codes, has begun to revolutionize predictive modeling of materials. Molecular dynamics (MD), in particular, has led to breakthrough advances in diverse fields, including tribology, energy storage, catalysis, sensing. Furthermore, recent integration of MD simulations with X-ray characterization has demonstrated promise in real-time 3-D atomistic characterization of materials. The popularity of MD is driven by its applicability at disparate length/time-scales, ranging from ab initio MD (hundreds of atoms and tens of picoseconds) to all-atom classical MD (millions of atoms and tens of nanoseconds), and coarse-grained (CG) models (microns and tens of micro-seconds). Nevertheless, a substantial gap persists between AIMD, which is highly accurate but restricted to extremely small sizes, and those based on classical force fields (atomistic and CG) with limited accuracy but access to larger length/time scales. The accuracy and predictive power of classical MD is dictated by the empirical force fields, and their capability to capture the relevant physics. In this talk, we will present some of our recent work on the use of machine learning (ML) to seamlessly bridge the electronic, atomistic and mesoscopic scales for materials modeling. Our automated ML framework aims to bridge the significant gulf that exists between the handful of research groups that develop new interatomic potential models (often requiring several years of effort) and the increasingly large user community from academia and industry that applies these models. Our ML approach has showed marked success in developing force fields for a wide range of materials from metals, oxides, nitrides, hetero-interfaces to two-dimensional (2-D) materials and even water (arguably the most difficult system to capture from a molecular perspective). This talk will also briefly discuss our ongoing efforts to integrate such cheap yet accurate atomistic models with (a) AI techniques to perform inverse design and construct metastable phase diagrams of materials (b) Deep learning to improve spatiotemporal resolutions of ultrafast X-ray imaging.

Developments in machine learning have opened the door to a new generation of methods for studying phase transitions, due to their ability to efficiently and autonomously identify complex patterns in many-body systems. Applications of machine learning techniques in statistical physics vary from locating phase transitions in spin systems to pinpointing weak spots in colloidal glasses. The rapid emergence of machine learning applications in both statistical mechanics and materials science demonstrates that these techniques are destined to become an important tool for soft matter physics.

One problem this tool seems ideally suited to solve, both experimentally and in simulations, is the identification of different (crystalline) phases – even on the level of a single particle. For example, when we use either experiments or simulations to look at crystallization at the microscopic level, it is always a challenge to determine which particles are part of the fluid, and which ones have already crystallized. In this talk I will describe how machine learning techniques can automatically figure this out for us, even if we do not know in advance what crystal structure to look for. We will also examine how such methods might reveal new structural insights in “purely” disordered systems like glasses.

Block copolymer (BCP) self-assembly, and in particular directed self-assembly (DSA), has provided a powerful route to fabricate complex thin film structures at small length scales. Despite DSAs success in fabricating various 2D patterns, fabrication of complex 3D nanostructures remains a challenge. Here, we will introduce a novel method, which we call Self-Directed Self-Assembly, and demonstrate how it can be leveraged to assemble 3D designer nanostructures using block copolymers. In particular we will demonstrate that uniform and non-uniform multilayer nanostructures can be obtained through stacking two or more "orthogonal" block copolymers. Finally, by introducing a patterned substrate to the first layer and performing multilayer stacking, we will demonstrate how to propagate information upwards in the film. Different complex multilayer structures will be shown, and at the end we will present an outlook on exciting new computational and artificial intelligence avenues of research in this field.