Fast atomic structure optimization with on-the-fly sparse Gaussian process potentials.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Atomic structure optimization

Background:

• Accurate atomic structure optimization is crucial for materials discovery.
• First-principles calculations are computationally intensive, limiting their use in large-scale simulations.
• Machine learning potentials (MLPs) offer a promising alternative for accelerating these calculations.

Purpose of the Study:

• To investigate the efficacy of on-the-fly machine learning potentials (MLPs) using sparse Gaussian process regression (SGPR) for rapid atomic structure optimization.
• To assess the acceleration achieved compared to traditional first-principles (FP) methods.
• To evaluate the transferability and suitability of MLPs for global optimization strategies.

Main Methods:

• Implementation of on-the-fly sparse Gaussian process regression (SGPR) for MLPs.
• Application to local optimization of atomic structures.
• Testing on random gold clusters to assess force reduction.
• Sequential optimization using MLPs for global structure searching.

Main Results:

• Significant acceleration in atomic structure optimization, even for single local optimizations.
• Forces reduced to approximately 0.1 eV/Å within ten first-principles (FP) calculations for random gold clusters.
• Demonstrated suitability of highly transferable MLPs for global optimization methods.
• Reduced reliance on FP calculations during sequential optimization of gold clusters.

Conclusions:

• On-the-fly MLPs with SGPR provide substantial speedups for atomic structure optimization.
• MLPs, despite potential accuracy limitations for exact local minima, are highly effective for accelerating global structure searches.
• This approach significantly reduces computational cost, making large-scale materials exploration more feasible.