Our paper “Geometry, Not Energy Surface, Drives the Neutral MLIP–DFT Gap in Atomistic Interaction Surrogates”, the study behind RAPIDS, has been accepted at the AI4Physics workshop at ICML 2026 (OpenReview).

A narrated walkthrough of RAPIDS — what the MLIP–DFT benchmark measures, and why geometry turns out to be the culprit.

RAPIDS is the rapid physical validation engine of our autonomous discovery stack. It benchmarks machine-learning interatomic potentials (MLIPs) against DFT across 5,567 probe–target dimer interactions and 18 benchmark tasks. The finding behind the title: geometric representation, not the energy surface itself, drives the neutral MLIP–DFT gap.

The packaging matters as much as the finding. RAPIDS is exposed as a tool autonomous LLM agents can call, so agents like DToR and screening twins like T³ can get a fast atomistic sanity check before committing to a costly simulation or experiment.

Co-first-authored with Zixin Ding and Rodrigo P. Ferreira, together with Yuxin Chen and Junhong Chen. I’ll present it in July.