Multifidelity Statistical Machine Learning for Molecular Crystal Structure Prediction.

Area of Science:

• Computational materials science
• Statistical machine learning
• Solid-state chemistry

Background:

• Predicting crystal structures requires accurate energies from first-principles calculations, often using density functional theory (DFT).
• High-throughput screening of numerous crystal structures is computationally prohibitive due to the cost of accurate DFT methods.
• Existing methods struggle to balance accuracy and computational expense for large-scale structure prediction.

Purpose of the Study:

• To develop a cost-effective machine learning approach for predicting accurate DFT energies of crystal structures.
• To enable efficient ranking and assessment of low-energy crystal structures.
• To reduce the computational burden of first-principles materials discovery.

Main Methods:

• A multifidelity machine learning approach using autoregressive Gaussian processes.
• Utilizing inexpensive force field and GGA DFT (PBE) calculations to bridge to expensive hybrid DFT (PBE0) energies.
• Benchmarking on crystal structure landscapes of three small, hydrogen-bonded organic molecules.

Main Results:

• Accurate prediction of PBE0 energies with errors < 1 kJ mol⁻¹.
• Achieved energy predictions at 4.2–6.8% of the cost of full PBE0 calculations.
• Demonstrated accurate prediction of crystal structure energies and ranking.

Conclusions:

• The developed probabilistic machine learning model accurately predicts crystal structure energies and rankings cost-effectively.
• This approach significantly reduces computational requirements for materials discovery.
• Understanding prediction uncertainties is crucial for reliable energetic ranking of crystal structures.