Benchmarking data efficiency in Δ-ML and multifidelity models for quantum chemistry.

Area of Science:

• Computational chemistry
• Machine learning applications
• Quantum mechanics

Background:

• Machine learning (ML) significantly reduces computational costs in quantum chemistry (QC).
• Generating training data for ML in QC remains a significant cost.
• Multifidelity machine learning (MFML) methods utilize data from multiple QC accuracy levels to mitigate costs.

Purpose of the Study:

• To compare the data generation costs of various ML methods for QC.
• To evaluate Δ-ML, MFML, optimized MFML, and a new MFΔML method.
• To assess prediction accuracy for ground state energies, excitation energies, and dipole moments.

Main Methods:

• Utilized the multifidelity benchmark dataset QeMFi.
• Compared data costs of Δ-ML, MFML, optimized MFML, and MFΔML.
• Benchmarked against single-fidelity kernel ridge regression.

Main Results:

• Multifidelity methods outperform standard Δ-ML for numerous predictions.
• The novel MFΔML method shows advantages for a limited number of predictions.
• MFΔML offers a cost-effective approach for specific ML applications in QC.

Conclusions:

• MFML strategies are superior to Δ-ML for large-scale QC predictions.
• The MFΔML method provides a valuable alternative for ML models requiring few evaluations.
• Optimizing data generation cost is crucial for efficient ML-driven QC.