Imitation and Accuracy in Automated Electronic-Structure Method Selection
Abstract
Scientific agents mine recorded computational choices, but those records do notshow whether another choice would be more accurate.We test this gap in electronicstructure method selection, where predictions can be executed against independent reference energies.Using 7,034 records mined from public repositories, we route requests by chemistry and compare BM25 retrieval, per-cell LoRA fine-tuning, and their combination in two evaluated cells.On the primary upload-grouped validation split, the backends are within three percentage points in exact match to mined choices.Relative to a training-mode constant, their margins are 11.7-14.4points for organic systems and at most 0.6 for the metal cell.A separate rebalanced run changes the organic backend ordering.Its raw metal margin of about ten points over a validationmode comparator falls to 5-6 points after 16 malformed gold labels are removed from 1 scoring.Because training and evaluation data also change, these runs do not isolate a split effect.We then execute predicted methods on benchmark reactions.Three fixed settings outperform the learned selectors in all nine tabulated organic comparisonssix distinct, since two constants coincide, and each scored only on the reactions both conditions could compute -while metal comparisons are mixed.A router trained on executed error improves over both members of one of four candidate pairs in development by 3.48 kcal/mol (95% CI 1.53-5.77),but does not beat fixed MP2 outside that pair.A frozen three-way router fails to improve on MP2 over 32 held-out reactions, and measured method-cost ordering reverses with system size.These findings specify checks for method-choice cells-label validity, within-cell variation, fixed comparators, target property, and cost-without establishing transferable routing.
Related
- VIBE-QC: an open-source, general-purpose electronic-structure code for molecules and solids developed with generative AI
- Electronic Descriptors Lower Oxygen-Vacancy Formation-Energy Error Relative to Crystal Structure Alone in Non-Magnetic Oxides
- Electronic Structure, Bonding Topology, and Photophysical Properties of the Low-Valent Beryllium Cluster CpBe2NO: A Multireference and Real-Space Wavefunction Analysis
- Synthetic Mechanism of a Fe(II) N -Heterocyclic Carbene Bidentate Complex Revealed by Electronic Structure Methods
- The Evolution of Electronic Structure in Conjugated Molecules Revealed by Mixed-Reference Spin-Flip Time-Dependent Density Functional Theory and Density-Based Analyses: II. Polyacenes
BibTeX
@misc{liu2026imitation,
title = {{Imitation and Accuracy in Automated Electronic-Structure Method Selection}},
author = {Qianying Liu and Yuki Kurashige and Ruiqi Zhang},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010313/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010313/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Imitation and Accuracy in Automated Electronic-Structure Method Selection
AU - Qianying Liu
AU - Yuki Kurashige
AU - Ruiqi Zhang
PY - 2026
DA - 2026/10/12
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010313/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010313/v1
AB - Scientific agents mine recorded computational choices, but those records do notshow whether another choice would be more accurate.We test this gap in electronicstructure method selection, where predictions can be executed against independent reference energies.Using 7,034 records mined from public repositories, we route requests by chemistry and compare BM25 retrieval, per-cell LoRA fine-tuning, and their combination in two evaluated cells.On the primary upload-grouped validation split, the backends are within three percentage points in exact match to mined choices.Relative to a training-mode constant, their margins are 11.7-14.4points for organic systems and at most 0.6 for the metal cell.A separate rebalanced run changes the organic backend ordering.Its raw metal margin of about ten points over a validationmode comparator falls to 5-6 points after 16 malformed gold labels are removed from 1 scoring.Because training and evaluation data also change, these runs do not isolate a split effect.We then execute predicted methods on benchmark reactions.Three fixed settings outperform the learned selectors in all nine tabulated organic comparisonssix distinct, since two constants coincide, and each scored only on the reactions both conditions could compute -while metal comparisons are mixed.A router trained on executed error improves over both members of one of four candidate pairs in development by 3.48 kcal/mol (95% CI 1.53-5.77),but does not beat fixed MP2 outside that pair.A frozen three-way router fails to improve on MP2 over 32 held-out reactions, and measured method-cost ordering reverses with system size.These findings specify checks for method-choice cells-label validity, within-cell variation, fixed comparators, target property, and cost-without establishing transferable routing.
ER -