Source-linked AI summary
Transition1x -- a Dataset for Building Generalizable Reactive Machine Learning Potentials
Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Jonas Busk, Ole Winther
TL;DR
Machine-learning potentials have had limited success predicting general reaction barriers, with training data from transition-state regions remaining scarce. Transition1x uses NEB-based pathway exploration to collect DFT calculations, and models trained on ANI1x alone lack sufficient transition-state data while Transition1x-trained models perform better on ANI1x than the reverse.
Problem
Accurate, general prediction of reaction barriers remains challenging because training datasets with extensive sampling of reactive and high-energy regions are lacking.
Source: §I · p. 1Method
Transition1x uses NEB-based exploration of reaction pathways to generate DFT energies and forces, retaining sufficiently changed intermediate paths.
Source: §I · p. 2, Nudged Elastic Band · p. 3, Data selection · p. 4Results
Models trained on ANI1x lack sufficient transition-state data, while Transition1x-trained models perform better on ANI1x than ANI1x-trained models do on Transition1x.
Source: Technical Validation · p. 10Takeaways & Limitations
Transition-state data is relevant to using ML potentials in NEB or GSM algorithms and training molecular-dynamics models for transition-state regions.
Source: Technical Validation · p. 11
Abstract
from arXiv · showhide
Machine Learning (ML) models have, in contrast to their usefulness in molecular dynamics studies, had limited success as surrogate potentials for reaction barrier search. It is due to the scarcity of training data in relevant transition state regions of chemical space. Currently, available datasets for training ML models on small molecular systems almost exclusively contain configurations at or near equilibrium. In this work, we present the dataset Transition1x containing 9.6 million Density Functional Theory (DFT) calculations of forces and energies of molecular configurations on and around reaction pathways at the wB97x/6-31G(d) level of theory. The data was generated by running Nudged Elastic Band (NEB) calculations with DFT on 10k reactions while saving intermediate calculations. We train state-of-the-art equivariant graph message-passing neural network models on Transition1x and cross-validate on the popular ANI1x and QM9 datasets. We show that ML models cannot learn features in transition-state regions solely by training on hitherto popular benchmark datasets. Transition1x is a new challenging benchmark that will provide an important step towards developing next-generation ML force fields that also work far away from equilibrium configurations and reactive systems.
I. Introduction
Reactive machine-learning potentials need training data from high-energy and transition-state regions, which popular datasets largely lack. Transition1x addresses this gap with NEB-based exploration of reaction pathways and DFT calculations, and the paper presents it as a resource for training models across diverse reaction surfaces.
I. Introduction
- Transition1x uses NEB-based exploration to generate millions of DFT calculations around reaction pathways, addressing the scarcity of reactive-region training data.The workflow starts from reactant–product pairs and saves intermediate calculations when paths move significantly.Source: §I · pp. 1–2, Methods · p. 2
- The source set contains 11,961 reactions based on GDB7, with up to seven heavy atoms including C, N, and O.The original reactants, products, and transition states were generated with the Growing String Method at ωB97X-D/def2-TZVP.Source: Initial Data · p. 2
- CINEB is applied after regular NEB reduces the maximal perpendicular force below 0.5 eVÅ−1, then continues until that force is at most 0.05 eVÅ−1.The thresholds balance path accuracy against redundant DFT calculations; transition states were not further refined because the goal was pathway-near configurations.Source: Nudged Elastic Band · p. 3
- 10073 reactions converged; nonconvergent reactions were discarded, and new paths were saved when cumulative Fmax since the last saved path exceeded 0.1 eVÅ−1.Early iterations often contributed every path, while additions became less frequent near convergence.Source: Data selection · p. 4
- PaiNN models are trained and evaluated on Transition1x, QM9x, and ANI1x, with force predictions derived from energy gradients.QM9x recalculates QM9 geometries at the appropriate DFT level for comparison with Transition1x.Source: Model and Training · p. 4, QM9 and QM9x · p. 4
ANI1x
The dataset comparisons show that Transition1x-trained models transfer to ANI1x better than ANI1x-trained models transfer to Transition1x, while QM9x-trained models perform poorly on both out-of-equilibrium datasets. The accompanying analyses contrast pathway convergence and transition-state agreement, and show that Transition1x samples distinct force and heavy-atom distance distributions.
ANI1x
- 10,073 of 11,961 reactions converged with NEB, and 89% of converged reactions did so within 200 iterations; nonconverged reactions were discarded.The authors report that nonconverged reactions often contain unphysical structures that are unsuitable as training data.Source: Technical Validation · p. 5
- NEB-found transition states correspond to original GSM transition states with MAE 0.16 eV, RMSE 0.19 eV, and average RMSD 0.11 Å; NEB barrier energies tend to be shifted higher.Source: Technical Validation · p. 7
- Transition1x's force distribution cusps at zero, whereas ANI1x has flatter, wider tails and higher force variance despite its geometries being closer to equilibrium.For hydrogen, Transition1x shows a sudden drop in the force distribution, which the authors relate to relaxation during data generation.Source: Technical Validation · pp. 7, 8, 10, Figure 5 · p. 8
- Transition1x covers more varied heavy-atom distances than ANI1x; NEB samples low-energy valleys containing likely molecules with otherwise energetically unfavorable distances.The authors note that such extreme distances are difficult to produce through ANI1x's normal-mode sampling technique.Source: Figure 6 · p. 9, Technical Validation · p. 10
- Transition1x-trained models perform better on ANI1x than ANI1x-trained models do on Transition1x; QM9x-trained models perform poorly on both Transition1x and ANI1x.The test sets exclude formulas seen during training; the authors attribute the asymmetry to ANI1x lacking sufficient transition-state data and QM9x containing only equilibrium structures.Source: Technical Validation · pp. 10–11, Table 1 · p. 10
Code availability
The analysis scripts are available upon request, and the passage also records the authors’ contributions.
Code availability
- Analysis scripts are available upon request; calculations used ORCA 5.0.2 and ASE 3.22.1 for electronic structure and NEB, respectively.Source: Code availability · p. 11
- M.S. led coding, experiments, and most of the writing, with coauthors contributing study design, supervision, review, and discussions.Source: Author contributions statement · p. 11
A. Comparison of GSM and NEB
Figure 7 compares transition states and barriers obtained with NEB and 6-31G(d) in this work against GSM and def2-mSVP in the original work.
A. Comparison of GSM and NEB
- Figure 7 compares NEB/6-31G(d) results with GSM/def2-mSVP results using a transition-state RMSD histogram and an energy plot.Panel (a) shows the RMSD distribution; panel (b) plots NEB energies on the x-axis against GSM energies on the y-axis.Source: Figure 7 · p. 15