Machine-Learning combined with Genetic Algorithms for Alloy Clusters Discovery

DFT-trained neural networks (NN) have been shown to dramatically reduce computational costs for predicting atomic properties without accuracy loss. However, this typically demands the DFT training data to be computed ahead of time. While this approach works for specialized applications, it is still infeasible for exploration into the vast configuration spaces inherent in nanoclusters of various sizes and compositions. To remedy this, we include in our methodology a Genetic Algorithm (GA) for generating structures or adsorption sites unbiasedly across configuration spaces. A self-optimizing program is developed to gradually train a single NN 'on-the-fly' for configurations generated by the GA and validated with DFT. When the self-optimization is complete, a NN capable of predicting nanocluster energies and forces for any reasonable structure or adsorption site within the configuration space is produced. This allows us to unbiasedly explore these spaces at the DFT level with relatively low computational demand automatically. The self-optimizer is tailored to explore the configuration or adsorption sites of a given nanocluster size, and composition, and can easily be extended to surface slab adsorption site exploration as well. We will present results for 13-atoms AuPd clusters.