Accelerating and scaling on-lattice Solid-On-Solid kinetic Monte Carlo simulations using an autoregressive Conditional Variational Autoencoder

Abstract

Kinetic Monte Carlo (kMC) simulations are fundamental for modeling atomistic surface morphology evolution, but their inherently sequential execution of atomic steps creates severe computational bottlenecks. This computational burden restricts accessible time-and length scales, making dynamic coupling with continuum solvers, such as Computational Fluid Dynamics, difficult. To overcome these limitations, this work introduces a novel surrogate model based on an autoregressive fully convolutional Conditional Variational Autoencoder (CVAE). As a proof of concept, we demonstrate the acceleration and scaleup of a 1D+1 Solid-on-Solid kMC model. By predicting the structural evolution over large time leaps and conditioning on thermodynamic parameters like temperature, the CVAE accurately captures the statistical surface growth and roughness dynamics across long time scales. The surrogate maintains high predictive fidelity while achieving massive computational speed-ups, bypassing expensive transition rate recalculations during temperature fluctuations, condensing many sequential events into a single forward pass, and demonstrating spatial scalability to significantly larger lattices. This approach provides a great speed-up for kMC simulations, paving the way for efficient multiscale simulation frameworks.