Beyond Candidate Generation: A Validation-First Protocol for AI-Assisted Transition-State Discovery
Abstract
Machine-learned models can now propose three-dimensional transition-state (TS) geometries in seconds, but a candidate geometry is not a validated TS. The gap between a plausible structure and a mechanistically meaningful barrier—one imaginary frequency, the intended atom pair, the correct connectivity, and a consistent electronic state—remains the bottleneck in AI-assisted computational chemistry. We present a validation-first protocol that separates candidate generation from physical validation. The protocol combines (i) a chemical preflight that fixes reactive pairs, spin surfaces, connectivity, and conformers; (ii) a per-point independent scan that localizes the barrier without fragile collective scans; (iii) a constrained reference calculation that produces the Hessian; (iv) an OptTS stage that reuses that Hessian; and (v) a validation stage based on imaginary-mode localization, key bond lengths, IRC endpoints, spin diagnostics, and energy-convention checks. The protocol is implemented as a portable, open-source skill package for local ORCA/xtb and Gaussian/Slurm execution. Case studies from Ni/Fe catalysis illustrate the failure modes the protocol is designed to catch: a generative model produced candidates with 3.6–4.4 Å heavy-atom RMSD for an 84-atom broken-symmetry organometallic system, a collective scan failed three times where a per-point strategy completed 16/16 points, and a single-conformer R/S comparison would have misread a conformerdominated difference. The reported gains of multi-sampling and conformer enumeration are taken from the published benchmark, not measured here. We argue that the near-term role of AI is not to replace electronic-structure judgment but to shrink the search space before it and to make the validation layer reproducible.
Related
- Cutoff-Safe Tabular Fusion of TESS Data-Validation Products for Temporal Candidate Vetting
- Transition magnetic-dipole dark matter and the LZ230616 high-recoil candidate
- Defect-Mediated Nucleation and Dynamics across the Phase Transition in the Excitonic Insulator Candidate Ta2NiSe5
- Necessary conditions for benchmarking next-generation AI in materials discovery
- Data-driven candidate structure ensemble generation beyond the convex hull with crystALL approach
BibTeX
@misc{li2026beyond,
title = {{Beyond Candidate Generation: A Validation-First Protocol for AI-Assisted Transition-State Discovery}},
author = {Haofan Li},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15011128/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011128/v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Beyond Candidate Generation: A Validation-First Protocol for AI-Assisted Transition-State Discovery
AU - Haofan Li
PY - 2026
DA - 2026/10/27
PB - ChemRxiv
DO - 10.26434/chemrxiv.15011128/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15011128/v1
AB - Machine-learned models can now propose three-dimensional transition-state (TS) geometries in seconds, but a candidate geometry is not a validated TS. The gap between a plausible structure and a mechanistically meaningful barrier—one imaginary frequency, the intended atom pair, the correct connectivity, and a consistent electronic state—remains the bottleneck in AI-assisted computational chemistry. We present a validation-first protocol that separates candidate generation from physical validation. The protocol combines (i) a chemical preflight that fixes reactive pairs, spin surfaces, connectivity, and conformers; (ii) a per-point independent scan that localizes the barrier without fragile collective scans; (iii) a constrained reference calculation that produces the Hessian; (iv) an OptTS stage that reuses that Hessian; and (v) a validation stage based on imaginary-mode localization, key bond lengths, IRC endpoints, spin diagnostics, and energy-convention checks. The protocol is implemented as a portable, open-source skill package for local ORCA/xtb and Gaussian/Slurm execution. Case studies from Ni/Fe catalysis illustrate the failure modes the protocol is designed to catch: a generative model produced candidates with 3.6–4.4 Å heavy-atom RMSD for an 84-atom broken-symmetry organometallic system, a collective scan failed three times where a per-point strategy completed 16/16 points, and a single-conformer R/S comparison would have misread a conformerdominated difference. The reported gains of multi-sampling and conformer enumeration are taken from the published benchmark, not measured here. We argue that the near-term role of AI is not to replace electronic-structure judgment but to shrink the search space before it and to make the validation layer reproducible.
ER -