Constructing surrogates for atomistic simulations via deep learning and generative large language models

Abstract Atomistic simulations offer insights into material behavior at the atomic level. However, they can be computationally intensive. In this paper, two deep learning models, deep symbolic optimization (DSO) and deep neural networks (DNN), and one generative large language model, GPT-4o, are employed to construct surrogates for atomistic simulations. Specifically, atomistic simulations are first performed to investigate the collapse of a nanovoid under hydrostatic pressure. We focus on the role of initial void radius and material characteristics, such as intrinsic and unstable stacking fault energies and surface energy (SE). We find that the critical pressure required for void collapse spans from 17.05 to 19.62 GPa, with the highest values corresponding to the maximum USFE. Additionally, an intermediate SE value (1068.13 $$\hbox {mJ/m}^2$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mtext>mJ/m</mml:mtext> <mml:mn>2</mml:mn> </mml:msup> </mml:math> ) minimizes the critical pressure. Based on the simulation results, surrogate models based on DSO, DNN, and GPT-4o are constructed, concluding that the SE affects the critical pressure the most. Graphical abstract