An approach based on genetic algorithms and machine learning coupled for studying alloy and molecular clusters by optimizing quantum energy surfaces

Abstract

A new genetic algorithm has been proposed focusing on direct ab initio potential energy surface (PES) global minima search. Besides the commonly used operators, this new approach uses an operator to: improve the initial cluster generation, classify and compare all generated clusters, and use machine learning to model the quantum PES used in parallel optimization. Part of the validation process for this methodology was done with $C{u}_{n}A{u}_m$ ( $n+m\le X$ for $X=14,19,38,55$ ) and $A{u}_{n}A{g}_n$ ( $n=10,20,30,40,50,60,70$ , and 75). The results are in fair agreement with the literature and led to a new global minimum for $C{u}_{12}A{u}_7$ . A search has been done for the lowest energies of $L{i}_n$ nanoclusters with 2-8 atoms using the DFT approach and for $L{i}_3,L{i}_4,L{i}_{2}H$ , $L{i}_{3}H$ using DLPNO-CCSD(T) approach. NQGA successfully performed the MP2 optimizations for ${\left(H_{2}O\right)}_{11}$ cluster. In all cases, the proposed genetic algorithm located the previously reported global minima with very efficient performance. The new proposed methodology makes it possible to optimize cluster geometries directly using high-level ab initio methods relinquishing any bias introduced by a classical approach. Our results show that this proposed method has great potential applications due to its flexibility and efficiency in identifying global minima in the tested atomic systems.

Keywords: ab initio potential energy surface; genetic algorithm; machine learning; metal clusters.