ChIMES Machine-Learned Interatomic Potential for Hydrogen Phase Diagram Prediction

Classical force fields, while computationally inexpensive, are often inadequate for materials discovery because their functional forms do not suitably describe conditions beyond equilibrium. Complex classical models are also difficult to fit, due to their non-linear form. Quantum methods, such as density functional theory (DFT), can provide an accurate model towards design and discovery, but in the case of large phase space or noncrystalline materials, can be prohibitively computationally expensive . Machine learning can bridge this gap; our machine-learned interatomic potential (MLIAP), ChIMES¹ can achieve quantum level accuracy, with significantly improved computational efficiency.<br/><br/> ChIMES is a physics-informed MLIAP, which describes system energy through an explicitly many-body expansion of fully-connected cluster graphs. It uses Chebyshev polynomials as the basis, yielding a parametrically linear model that can be rapidly generated and has the unique feature of being explicitly chemically extensible Here we highlight that ChIMES can model hydrogen in the region between 0 and 100 GPa, for 300-10,000 K. Understanding this region is crucial for designing and discovering materials that undergo shock, and are subject to extreme conditions. Yet we know very little about the corresponding hydrogen phase diagram due to both experimental and computational limitations. Experimentally, techniques used to drive materials to the relevant conditions preclude measuring temperature and characterizing structural and chemical transformations.² In this case study, regularization was through the Lasso algorithm, and penalty functions were used to tune the distance of interatomic interactions. The polynomial order was also tuned. Moreover, ChIMES has the capacity for active learning ³, so that the model improves iteratively. Active learning makes materials discovery a more targeted approach: ChIMES can learn how to improve the training database. The broader goal is to transfer this honed model to any C,H,O,N system using ChIMES.<br/><br/> The ability of ChIMES to simulate overtimes beyond the DFT simulation time is key to understanding phase changes, especially those mediated by chemistry. The ChIMES algorithm maintains force, energy and stress permutational, translational and rotational invariance, enabling facile thermodynamic analysis. For hydrogen, simulating trajectories for the unknown region of this phase diagram is useful for material discovery. Under high enough temperature and pressure, hydrogen can dissociate. This chemistry dictates whether spin polarization should be included in calculations. However, the loci for this transition was not known <i>a priori. </i>Hence, ChIMES enabled the exploration and discovery of this transition.<br/><br/> Ultimately, ChIMES can serve as a key enabling tool for increasing materials discovery workflow efficiency and accuracy. In this case, rather than modeling hydrogen in conventional quantum methods by assembling a collection of DFT trajectories that likely will not cover the region’s phase diagram, the ChIMES MLIAP allows exploration of states beyond the training set.<br/><br/>1.R.K. Lindsey*, L.E. Fried, N. Goldman, <i>J. Chem. Theory Comput.</i> <b>13</b> 6222 (2017)<br/>2.Nellis, W. J. "Dynamic compression of materials: metallization of fluid hydrogen at high pressures." <i>Reports on progress in physics</i> 69, no. 5 (2006): 1479<br/>3. R.K. Lindsey, L.E. Fried, N. Goldman, S. Bastea, <i>JCP</i>, <b>153</b> 134117 (2020)