Accelerating Crystal Structure Prediction of Complex Organic Molecules Using Machine Learning Interatomic Potentials and Simulated Annealing

Crystal structure prediction (CSP) is a longstanding challenge in materials science, with conventional methods relying on computationally expensive density functional theory (DFT) calculations that limit their applicability to small systems. Recent advances in machine learning interatomic potentials (MLIPs) offer a paradigm shift, enabling efficient exploration of vast configuration spaces with near-DFT accuracy at greatly reduced cost. While MLIPs have shown promise in accelerating CSP, existing approaches continue to face challenges in predicting ground-state structures of complex organic molecules. Here, we introduce an MLIP-driven active-learning CSP workflow that integrates simulated annealing molecular dynamics with DFT, enabling efficient and accurate prediction of stable crystal structures of complex organic molecules. We demonstrate our workflow on polyamine molecular systems, and we show that our workflow efficiently generates low-energy conformations of complex polyamines while simultaneously yielding accurate MLIPs and high-fidelity DFT reference data. The resulting database can be integrated with external datasets to improve MLIP transferability across diverse materials systems.

Acknowledgements

Calculations are performed using NERSC.