How accurate are DFT forces? Unexpectedly large uncertainties in molecular datasets

Area of Science:

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

Background:

• Machine learning interatomic potentials (MLIPs) are increasingly used for materials simulations.
• Training MLIPs requires large datasets, typically generated using density functional theory (DFT).
• Accuracy of MLIPs depends critically on the quality and convergence of the underlying DFT data.

Purpose of the Study:

• To assess the quality of DFT-generated data used for training MLIPs.
• To identify potential sources of error in commonly used DFT datasets.
• To quantify the impact of DFT inaccuracies on force components.

Main Methods:

• Analysis of net forces in several prominent DFT datasets (SPICE, Transition1x, ANI-1x, ANI-1xbb, AIMNet2, QCML, OMol25).
• Recalculation of forces using more converged DFT settings for error quantification.
• Comparison of force components between original and recomputed datasets.

Main Results:

• Several analyzed DFT datasets exhibit significant nonzero net forces, indicating potential convergence issues.
• Individual force component errors were quantified, with average discrepancies ranging from 1.7 meV/Å (SPICE) to 33.2 meV/Å (ANI-1x).
• Substantial discrepancies in force components were observed across multiple datasets.

Conclusions:

• The quality of DFT data is a critical bottleneck for developing accurate MLIPs.
• Non-converged DFT calculations introduce significant errors in force components, compromising MLIP reliability.
• Ensuring well-converged DFT data is essential for advancing MLIP development and applications.