Finite-Temperature Crystal Structure Prediction with Universal Neural Network Potential and Free Energy Calculation

Crystal structure prediction (CSP) is a critical process for predicting stable crystal structures in given systems, serving as a crucial prerequisite for harnessing computational atomic simulations. Typically, CSP methods are combined with density functional theory (DFT) calculations to evaluate the formation energy of the candidate structures. However, the efficiency of CSP is greatly hindered by the time-consuming DFT calculations, which limit the exploration of numerous candidate structures.<br/>Recent advancements of neural network potential (NNP) fitted by DFT calculations, offer an efficient approach to CSP thanks to their fast and accurate energy evaluation. Given the interest in multi-component systems in CSP, it is preferable for NNP to exhibit universality across various systems and a strict level of accuracy to capture subtle energy differences in distinct structures. We employed our developed universal NNP called PFP, which was trained with extensive datasets comprising around 59 million structures [1]. The PFP showed 4.2 meV/atom (MAE) in accuracy of the energy above the convex hull among near-the-hull structures up to 10 meV/atom in Materials Project's unary, binary, and ternary systems. We use PFP and a genetic algorithm (GA) based CSP method inspired from the sophisticated multi-objective optimization algorithms [2].<br/>We further extend our analysis to finite temperature as conventional CSP only considers enthalpy at zero temperature. This involves free energy calculation for candidate structures found in CSP. Although the free energy calculations are computationally intensive and require complex workflows, recent developments have led to automated and robust frameworks [3,4]. Similar to these previous works, we employ thermodynamic integration to compute Gibbs free energy and manage the complex workflows with Argo Workflows [5], ensuring robustness and scalability in handling diverse structures in CSP.<br/>For validation, we computed thermal properties using PFP and compared them with empirical potential results [6]. Additionally, we analyzed temperature-dependence of phase stability for near-the-hull structures in CSP search. Our work, combining the universal NNP, GA-based CSP method, and automated free energy calculations, should demonstrate the promise of these methods in materials discovery.<br/><br/>[1] S. Takamoto, et al., Nat Commun 13, 2991 (2022).<br/>[2] K. Deb, et al., IEEE Transactions on Evolutionary Computation 18, 577 (2014).<br/>[3] S. Menon, et al., Phys. Rev. Mater. 5, 103801 (2021).<br/>[4] L. Zhang, et al., Phys. Rev. Lett. 126, 236001 (2021).<br/>[5] https://github.com/argoproj/argo-workflows<br/>[6] S. Ryu, et al., Model. Simul. Mat. Sci. Eng. 16, 085005 (2008).