Global Optimization Techniques for Cluster Geometry

The determination of the lowest‐energy configurations of atomic and molecular clusters poses a formidable challenge due to the exponential growth of possible arrangements with cluster size. The underlying potential energy surface of a cluster is characterised by a multitude of local minima separated by energy barriers, demanding robust global optimisation strategies capable of escaping traps and efficiently sampling high‐dimensional landscapes. Classical approaches such as genetic algorithms and basin‐hopping exploit stochastic or heuristic moves to traverse the landscape, while more recent developments harness machine learning to guide the search. Bayesian optimisation frameworks build surrogate models of the energy surface to propose promising trial structures with minimal computational cost. Reinforcement learning paradigms treat structure generation as a sequential decision problem, rewarding algorithms that propose low‐energy isomers. The synergy of these methods with first‐principles calculations, notably density functional theory, has enabled routine exploration of clusters comprising tens to hundreds of atoms. Advances in data‐driven repositories of computed cluster geometries have in turn accelerated the development and benchmarking of new algorithms, paving the way for rational design of catalysts, nanoalloys and functional nanomaterials.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

A comprehensive open database of over sixty thousand low‐energy cluster structures across more than fifty elements has been released, providing a rich resource for both training machine‐learning potentials and benchmarking optimisation algorithms. This repository reveals systematic size‐dependent structural motifs and deepens understanding of chemical trends at the nanoscale. In parallel, a novel reinforcement learning framework has been introduced to generate low‐energy isomers of metal clusters, employing an actor–critic architecture and a learned interatomic potential as a reward function. This approach markedly reduces computational expense relative to conventional density‐functional‐theory‐based searches and uncovers previously unknown isomer families. Complementing these developments, active‐learning Bayesian optimisation has been adapted to the search for low‐energy conformers of organic molecules adsorbed on metallic clusters. By constructing a surrogate energy model and incorporating constraints to avoid steric clashes, this method achieves rapid convergence to global minima in high‐dimensional spaces, demonstrating its potential for in silico screening of cluster–ligand assemblies.

Global Optimization Techniques for Cluster Geometry publication trend

The graph below shows the total number of articles in global optimization techniques for cluster geometry across all publications each year (not limited to Nature Index journals).

Technical terms

Potential energy surface (PES): A multidimensional landscape representing the energy of a system as a function of atomic positions.

Global optimisation: The process of finding the absolute minimum of the PES among many local minima.

Genetic algorithm: A heuristic search method inspired by natural selection, employing crossover and mutation to evolve candidate structures.

Basin‐hopping: A technique that transforms the PES into a staircase of local minima, enabling facile exploration by random perturbations and minimisations.

Bayesian optimisation: A sequential strategy that uses a surrogate model of the PES to identify the most promising points to evaluate next.

Reinforcement learning: A framework in which an agent generates structures through a series of decisions, receiving rewards based on energy evaluations.

Density functional theory (DFT): A quantum mechanical method for computing electronic energy and forces in materials with a favourable balance of accuracy and cost.

Isomer: A distinct structural arrangement of the same chemical composition sharing identical stoichiometry but differing in geometric connectivity or spatial configuration.

References

1. A database of low-energy atomically precise nanoclusters. Scientific Data (2023).
2. Exploring the Conformers of an Organic Molecule on a Metal Cluster with Bayesian Optimization. Journal of Chemical Information and Modeling (2023).
3. MeGen - generation of gallium metal clusters using reinforcement learning. Machine Learning: Science and Technology (2023).