Accelerating Materials Discovery Through Sparse Gaussian Process Machine Learning Potentials

ConspectusQuantum mechanical calculations predict atomic behavior with high accuracy but scale cubically, limiting routine applications to hundreds of atoms. This computational bottleneck prevents materials discovery at realistic scales where complex chemistry occurs. Designing next-generation batteries, solar cells, and catalysts requires predicting atomic behavior under diverse conditions, yet direct quantum simulations remain prohibitively expensive for such systems.Machine learning potentials offer a transformative approach by learning from limited quantum calculations to enable simulations that are orders of magnitude faster. This Account presents sparse Gaussian process regression (SGPR) as a statistically rigorous framework that addresses two critical challenges: achieving quantum-level accuracy from minimal training data and providing calibrated uncertainty estimates essential for exploring unknown chemical spaces. Unlike state-of-the-art graph neural network approaches that typically require 10,000-100,000+ training structures, while enabling simulations of up to ∼10⁴-10⁵ atoms, SGPR achieves practical accuracy with merely 100-1000 quantum calculations and a 10-100× reduction in data requirements, though more suited for focused chemical spaces (for systems < ∼3000 atoms) rather than broad multielement coverage.The SGPR framework exploits redundancy in chemical environments through intelligent rank reduction, selecting only the most informative local chemical environments as inducing points. On-the-fly adaptive sampling triggers new quantum calculations only when model uncertainty exceeds defined thresholds, ensuring efficient data generation. The robust Bayesian committee machine (RBCM) architecture partitions large models into specialized expert models for specific chemical compositions, then combines them through confidence-weighted schemes. This modular approach achieves O(nm²/p²) scaling (for n data points, m inducing sets, and p experts) and enables systematic expansion to complex multicomponent systems without retraining entire models.Applications demonstrate the framework's versatility across diverse materials challenges. For solid electrolytes, SGPR potentials trained on ∼100 selected structures achieved force prediction accuracy sufficient for reliable molecular dynamics, enabling simulations of Li₇P₃S₁₁ supercells with 10⁴× speedup over direct DFT. The calculations revealed high intrinsic conductivities in the tens of mS/cm range with activation energies consistent with NMR measurements. These simulated single-crystal values are around the upper limit of typical experimental measurements on polycrystalline samples, likely reflecting the absence of grain boundary resistance in idealized simulations. For perovskite solar cells achieving ∼26% efficiency, SGPR-MD simulations revealed that SnO₂/perovskite interfaces spontaneously form stabilizing FASnCl₃ interlayers─mechanistic insights requiring extensive configurational sampling beyond static DFT capabilities. In electrocatalysis, screening identified Pt-C₂N₂ single-atom catalysts with overpotentials of ∼30 mV for chlorine evolution, significantly outperforming industrial catalysts (∼100 mV). For organic systems, unified potentials for hydrocarbons achieved subkcal/mol accuracy for π-π stacking interactions and correctly predicted polymer folding dynamics where empirical force fields systematically fail.The SGPR-RBCM framework offers distinct advantages when training data is limited, uncertainty quantification is critical, and focused chemical accuracy is prioritized. By enabling quantum-accurate simulations at near-classical computational costs, these methods accelerate high-throughput screening across diverse materials spaces. The modular expert architecture creates pathways toward comprehensive potentials that could transform materials discovery for clean energy, electronics, and sustainable chemistry applications.