Δ-learning for transferable machine learning interatomic potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine-learning interatomic potentials (MLIPs) often struggle with transferability and accuracy for out-of-distribution molecular geometries.
• Directly learning total energies can lead to unphysical behavior and significant errors, especially for transition states and conformers.

Purpose of the Study:

• To evaluate delta-learning (Δ-learning) as a method to enhance the accuracy and transferability of MLIPs.
• To develop and assess a Δ-learning model that corrects a third-order tight-binding density functional theory (DFTB3) baseline with a high-dimensional neural network potential (HDNNP).

Main Methods:

• Trained an ANI-style HDNNP as a correction to predict PBE0/aug-cc-pVTZ energies from a DFTB3 baseline.
• Evaluated the Δ-learning model (ANIDFTB-Δ) against a direct HDNNP model (ANIPBE0-Direct) on held-out test sets and benchmark datasets (GMTKN55, PX13, DES370K).

Main Results:

• The ANIDFTB-Δ model achieved a mean absolute error (MAE) of 0.82 kcal/mol, significantly outperforming the direct model (MAE of 2.01 kcal/mol).
• Δ-learning improved relative energies for conformers/tautomers, reduced errors in proton-transfer transition states, and corrected long-range electrostatic interactions.
• The DFTB3 baseline improved the description of intermolecular interactions and long-range electrostatics, reducing MAE from 0.098 to 0.019 kcal/mol.

Conclusions:

• Δ-learning offers a robust strategy to enhance MLIP accuracy and reliability, particularly for challenging chemical systems.
• The ANIDFTB-Δ model provides a practical balance between accuracy and computational cost for simulating moderately sized systems.
• Incorporating a physically informed baseline like DFTB3 significantly improves the performance of MLIPs, especially for properties sensitive to long-range interactions.