Train a reactive MLIP on Transition1x and predict reaction barrier heights to within 2 kcal/mol

Most ML interatomic potentials are trained only on near-equilibrium configurations and fail in transition-state regions. Transition1x supplies 9.6 million DFT (wB97x/6-31G(d)) energies and forces sampled ON and AROUND the reaction paths of about 10000 organic reactions, generated by saving intermediate Nudged Elastic Band images. Using the dataset's official train/validation/test split by reaction, train a machine-learning interatomic potential and evaluate on the held-out test reactions two things: (1) the per-atom force MAE and per-configuration energy MAE over all test configurations, and (2) the barrier-height error, i.e. the MAE between the model-predicted and the DFT reference energy difference from reactant to the transition-state (saddle) image, in kcal/mol. Standard equilibrium-trained potentials degrade sharply near saddle points; can an agent-built model trained on Transition1x predict barrier heights to within 2 kcal/mol MAE on unseen reactions?

Acceptance. FULLY RESOLVES: an MLIP trained on the Transition1x training reactions and evaluated on the official held-out test reactions, achieving a reactant-to-saddle barrier-height MAE $\le 2$ kcal/mol, reported together with the overall test force MAE (meV/$\text{\AA}$) and energy MAE (meV), the exact split used, model/hyperparameters, and a runnable training+evaluation script with weights. PARTIAL: any reproducible Transition1x-trained model reporting test force/energy MAE and barrier-height MAE (even above 2 kcal/mol) with full protocol disclosure, or a documented negative result quantifying how a named architecture degrades from equilibrium to transition-state regions. Metrics: force MAE (meV/$\text{\AA}$), energy MAE (meV), and barrier-height MAE (kcal/mol) on the test reactions.