New benchmark: xTB pretraining + Δ-learning wins 31/32 for neural interatomic potentials
bravo_abad · x · 2026-11-18
Simulating how a chemical reaction proceeds requires searching atomic rearrangements from reactants to products, guided by energies and forces from ML models. A new benchmark compared four ways to train neural interatomic potentials when DFT labels are expensive.
- Best recipe: pretrain on 17.6M cheap GFN2-xTB calculations, then learn only the gap to DFT (pretraining + Δ-learning).
- With the full DFT set, it achieved the lowest MAE for energies/forces in 31 of 32 comparisons across four architectures, and the most accurate reaction energies and barriers on a held-out set.
- Counterintuitive finding: better ML predictions don't guarantee better reaction searches — the authors also ran full reaction searches for at least one architecture, suggesting prediction accuracy and search success should be evaluated separately.
More from Research
- Researcher laments reviewer fatigue as coding agents flood academia with incremental papers — rishabh16_ · 2026-11-18
- GPBam: New German Public Law Benchmark Finds Retrieval Gains Limited, Essay Quality Tracks Statutory Knowledge — aaire · 2026-11-18
- AI agents uncover first known genome-wide frameshifting in bacteria — GaryMarcus · 2026-11-18
- Google launches Antigravity Science plugin, gives 10,000 scientists free top-tier access — DynamicWebPaige · 2026-11-18
- AI predicts viral evolution and designs antibodies that neutralize forecasted escape variants — BrianHie · 2026-11-18
- Atomistic foundation models hit data-quality wall: fit error now below DFT label error — matt_slotnick · 2026-11-18