Robust Machine Learned Inter-atomic Potentials for Molecular Dynamics Simulations

Molecular Dynamics (MD) simulations accelerated material discovery by performing property prediction at scales higher than Density functional theory (DFT). To make property prediction from MD simulations more robust, we propose an uncertainty quantification framework. The developed framework considers uncertainty in the parameters of the interatomic potential (IP) used to compute energy/forces between a set of atoms. We considered two distinct potentials namely AIREBO and Embedded atom method. The material of choice are graphene and aluminium. Sensitivity analysis was performed to understand the effects of IP parameter change on the final properties of interest. The sensitivity analysis showed that by varying the potential parameters as small as by 1%, the variance in the resulting mechanical properties was as high as 1e5. Basic properties such as elastic constant of graphene showed variation of 66% by changing AIREBO parameters only by 0.5%. Based on the sensitivity analysis results, a new robust version of each IP (AIREBO/EAM) was developed using Bayesian methodologies. Using samples from the resulting posterior distribution, MD simulations were performed and error bars for each property was obtained. The variance in the properties was within experimental tolerance exhibiting the reduction in sensitivity for both the material systems. It can be challenging to obtain a single classical IP to describe the interactions involved in multi-element systems. Various machine learning based potentials have been proposed to overcome this limitation however their computational cost can be high. The proposed model approximates the energy/forces using a linear combination of random features at a lower computational cost. This study provides results for three classes of materials, namely two-dimensional materials, metals, and semiconductors. Force and energy predictions made using the proposed method are in close agreement with density functional theory calculations, with training time that is 96% lower than standard kernel models. This study also develops IPs for High Entropy alloys based on random features. MD simulations for mechanical and thermal properties are within 7% of the corresponding DFT values. Complex interactions present at high temperatures were also correctly identified by the proposed IP.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.