Data-Efficient Multifidelity Training for High-Fidelity Machine Learning Interatomic Potentials.

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• Machine learning interatomic potentials (MLIPs) estimate potential energy surfaces (PES) from ab initio calculations, offering near-quantum accuracy at lower computational cost.
• High-fidelity databases are crucial for MLIP accuracy but are expensive to create, limiting their application to systems needing high chemical accuracy.

Purpose of the Study:

• To develop an MLIP framework capable of simultaneous training on multifidelity databases.
• To enable accurate learning of high-fidelity PES using minimal high-fidelity data by leveraging lower-fidelity data.

Main Methods:

• Utilized an equivariant graph neural network for the MLIP framework.
• Employed a multifidelity training approach using generalized gradient approximation (GGA) as low-fidelity and meta-GGA as high-fidelity data.
• Tested the framework on Li₆PS₅Cl and InGa_1-N systems.

Main Results:

• Achieved excellent accuracy with only 10% high-fidelity data compared to the low-fidelity set.
• Demonstrated high accuracy in Li-ion conductivity predictions (within 10% error) and InGa_1-N mixing energy (R² of 0.98).
• Showed that low-fidelity GGA data effectively infers information from uncovered high-fidelity spaces, improving accuracy and molecular dynamics stability.

Conclusions:

• The multifidelity learning framework significantly enhances MLIP performance for high-accuracy tasks, outperforming transfer learning and Δ-learning.
• The methodology is versatile, applicable to various systems and can be extended to higher fidelity levels, including coupled-cluster.
• This approach promises the development of highly accurate, bespoke, or universal MLIPs by efficiently expanding high-fidelity datasets.