Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials
- 1Department of Chemistry
- 2Colorado State University
Abstract
Selectivity in chemical reactions is often dictated by small energetic differences between competing transition states of less than 2 kcal mol -1 . Computational predictions of selectivity for organic and organometallic reactions must reach these levels of quantitative accuracy, while simultaneously navigating the mechanistic and conformational complexity involved. Here, we assess the suitability of foundation machine-learned interatomic potentials (MLIPs) for this task. We introduce the SELECT50 dataset, a chemically diverse collection of paired selectivity-determining transition state structures, evaluated at the ωB97M-V/def2-TZVPD//ωB97X-D3(BJ)/def2-TZVP level of theory, spanning representative cases of kinetically-controlled stereo-, chemo- and regioselectivities, along with competing mechanistic pathways. We compare the performance of four MLIPs, alongside semiempirical and low-cost quantum chemical approaches, in determining transition state geometries, energetic discrimination, and pathway ordering. The Universal Model for Atoms MLIPs UMA-M-1.1 and UMA-S-1.2 predict relative transition state energies with sub-kcal mol -1 accuracy and mean absolute errors of 0.36 and 0.44 kcal mol -1 , providing saddle points that closely reproduce triple-ζ DFT geometries. DFT single-point refinement further reduced the mean absolute error (MAE) to 0.25 kcal mol -1 , supporting DFT//MLIP as a practical strategy for selectivity prediction. The major reaction pathway was correctly assigned in up to 98% of cases. Beyond individual transition structures, MLIPs reproduce complete organic and organometallic reaction landscapes and enable full-system treatment of selectivity-determining assemblies containing up to 211 atoms. Finally, MLIP-derived transition state energetics strongly track experimental selectivity across Ni- and Pd catalyzed reaction families (R 2 = 0.87–0.90; ρ = 0.80), and with calibration, yield mean held-out MAEs of 0.12 and 0.63 kcal mol -1 . Our results position foundation MLIPs as practical surrogates for the exploration and screening stages of DFT-based mechanistic analyses, enabling the throughput required for large-scale catalyst, substrate, and reaction screening.
Related
- Computational insights into the electrostatic origins of mechanistic selectivity: human topoisomerase IIα enforces ATP hydrolysis via substrate-assisted catalysis.
- Selectivity in Homolytic Substitution at a Tertiary Phosphine Center: Mechanistic Insights into Dealkanation Reaction.
- From Electron Density to Ionic Conductivity for the Mechanistic Understanding of a Fast Charging Acetone Electrolyte
- Accelerating Proton Affinity Prediction with Multi-Fidelity Machine Learning
- Accelerating Phase Prediction via CALPHAD-Informed Machine Learning and Data Augmentation
BibTeX
@misc{manoj2026accelerating,
title = {{Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials}},
author = {Niket Manoj and Austin Pounder and Niko Ahmed and Abhijeet S. Bhadauria and Robert S. Paton},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010828/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010828/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Accelerating Mechanistic Understanding and Prediction of Chemical Selectivity with Foundation Machine-Learned Interatomic Potentials
AU - Niket Manoj
AU - Austin Pounder
AU - Niko Ahmed
AU - Abhijeet S. Bhadauria
AU - Robert S. Paton
PY - 2026
DA - 2026/10/22
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010828/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010828/v1
AB - Selectivity in chemical reactions is often dictated by small energetic differences between competing transition states of less than 2 kcal mol -1 . Computational predictions of selectivity for organic and organometallic reactions must reach these levels of quantitative accuracy, while simultaneously navigating the mechanistic and conformational complexity involved. Here, we assess the suitability of foundation machine-learned interatomic potentials (MLIPs) for this task. We introduce the SELECT50 dataset, a chemically diverse collection of paired selectivity-determining transition state structures, evaluated at the ωB97M-V/def2-TZVPD//ωB97X-D3(BJ)/def2-TZVP level of theory, spanning representative cases of kinetically-controlled stereo-, chemo- and regioselectivities, along with competing mechanistic pathways. We compare the performance of four MLIPs, alongside semiempirical and low-cost quantum chemical approaches, in determining transition state geometries, energetic discrimination, and pathway ordering. The Universal Model for Atoms MLIPs UMA-M-1.1 and UMA-S-1.2 predict relative transition state energies with sub-kcal mol -1 accuracy and mean absolute errors of 0.36 and 0.44 kcal mol -1 , providing saddle points that closely reproduce triple-ζ DFT geometries. DFT single-point refinement further reduced the mean absolute error (MAE) to 0.25 kcal mol -1 , supporting DFT//MLIP as a practical strategy for selectivity prediction. The major reaction pathway was correctly assigned in up to 98% of cases. Beyond individual transition structures, MLIPs reproduce complete organic and organometallic reaction landscapes and enable full-system treatment of selectivity-determining assemblies containing up to 211 atoms. Finally, MLIP-derived transition state energetics strongly track experimental selectivity across Ni- and Pd catalyzed reaction families (R 2 = 0.87–0.90; ρ = 0.80), and with calibration, yield mean held-out MAEs of 0.12 and 0.63 kcal mol -1 . Our results position foundation MLIPs as practical surrogates for the exploration and screening stages of DFT-based mechanistic analyses, enabling the throughput required for large-scale catalyst, substrate, and reaction screening.
ER -