Atomistic structure search using local surrogate model.

Area of Science:

• Computational materials science
• Atomistic simulations
• Machine learning in chemistry

Background:

• Global structure search methods are computationally intensive.
• Accurate potential energy surface calculations are crucial for materials discovery.
• Efficient exploration of complex energy landscapes is a key challenge.

Purpose of the Study:

• To develop and validate a local surrogate model for accelerating atomistic structure searches.
• To integrate the surrogate model with global structure search algorithms.
• To demonstrate the model's applicability across diverse atomistic systems.

Main Methods:

• Utilizing the Gaussian approximation potential (GAP) formalism.
• Employing the smooth overlap of atomic positions (SOAP) descriptor.
• Implementing mini-batch k-means for sparsification of local environments.
• Integrating the model into the Atomistic Global Optimization X (AG OMX) framework.
• Using the surrogate model to partially replace local relaxations in basin hopping searches.

Main Results:

• The local surrogate model demonstrates robustness across various systems: molecules, nanoparticles, surface-supported clusters, and thin films.
• Significant acceleration of structure search processes is achieved.
• The model facilitates transfer learning from smaller systems to larger ones.
• Concurrent searches across multiple stoichiometries become feasible.

Conclusions:

• The developed local surrogate model effectively accelerates global structure search methods.
• This approach offers a computationally efficient alternative for exploring materials' potential energy landscapes.
• The model's versatility and transfer learning capabilities present significant advantages for materials discovery and design.