Guided Adaptive Diffusion: An Evolutionary Framework for Multimodal Atomistic Structure Prediction

Abstract

Atomistic structure prediction requires search algorithms capable of locating global and local minima on high-dimensional, multimodal potential energy surfaces. Traditional algorithms tend to become less effective as the dimensionality of the search space increases. In this work, we introduce an adaptive diffusion framework that reinterprets the neural-network-based denoising process as an evolutionary search mechanism for structure optimization. The framework incorporates two key optimization mechanisms. First, geometric constraints provide physics-informed guidance during sampling. Second, a memetic approach combines the global diverse sampling capabilities of diffusion models with local gradient-based relaxation. Unlike heuristic evolutionary algorithms, which rely on predefined analytical update rules for comparatively simple search distributions, neural-network-based denoising learns the underlying structure of the search space directly from the full accumulated history of sampled configurations, enabling the representation of highly complex distributions. We benchmark the algorithm using Lennard-Jones and gold clusters, demonstrating its ability to locate the global minimum and an ensemble of low-energy local minima within a single evolutionary run. The results indicate that the algorithm remains effective on high-dimensional potential energy surfaces, maintaining both population diversity and search efficiency throughout the optimization.