Experimental Inter-Report Scatter Bounds Machine-Learning Accuracy for TADF Singlet–Triplet Gaps—Structure and Semi-Empirical Features Reach the Same Limit
4 affiliations
- 1Department of Physics, Faculty of Science
- 2University of Yaoundé
- 3University of Douala
- 4Faculty of Medicine and Health Sciences, Université de Sherbrooke, 3001, 12th Avenue Nord, Sherbrooke, QC, Canada, J1H 5N4
Abstract
The singlet–triplet gap ∆ESTis the screening parameter for thermally activated delayed fluorescence (TADF), yet independent experimental reports of the same compound scatter by 0.072 eV RMS, and the per-molecule median regressed on is known only to 0.049 eV, about half of any model’s error. This label noise sets a ceiling on every predictor, yet the models sit below it: neither noise nor capacity alone accounts for the remaining error. Against that ceiling, two-dimensional structure predicts as accurately as a semiempirical excited-state calculation: on 231 donor–acceptor molecules (212 Bemis–Murcko scaffolds, scaffold cross-validation), random forests from Morgan fingerprints (MAE 0.091 eV) and RDKit descriptors (0.100 eV) match natural-transition-orbital descriptors from a GFN2-xTB/sTDA excited state (0.096 eV) to within 0.017 eV on paired folds, so the excited-state step buys no accuracy. Two further results are specific to this corpus. First, under a source-paper split the accuracy advantage over a mean-value baseline disappears, but the excited-state descriptors keep the rank correlation the fingerprints lose (paired ∆ρ = 0.19 on a single split, 95 % CI narrowly excluding zero): they order candidates across laboratories without predicting the gap. Second, the leakage that inflates many computed-gap models is learner-dependent—given the S1 and T1 energies as features, a linear model returns ∆EST ≡ ES1 − ET1 exactly (R2 = 1.000) while an axis-aligned forest on the same features recovers only R2 = 0.59 and clears cross-validation unremarked. Progress past the ≈ 0.1eV ceiling needs a curated, solvent-controlled, single-source experimental benchmark—one that would also expose the remaining modelling headroom. Data, models, and code are available (Zenodo DOI: 10.5281/zenodo.17436069).
Related
- Using structure to evaluate the accuracy and transferability of machine learning potentials of biomolecules
- ML26: New Ingredients for an Integral-Features Density Functional with Broad Chemical Accuracy
- Accuracy, Reliability, and Efficiency of a MACE Machine-Learning Interatomic Potential for Aspirin: A Reproducible Benchmark for Molecular Simulation
- A Molecular Generation–Validation Workflow Combining Reversible Junction Tree Reinforcement Learning and Semi-empirical Quantum Chemistry: A Case Study on Molecular Solar Thermal Fuels
- Imitation and Accuracy in Automated Electronic-Structure Method Selection
BibTeX
@misc{njafa2026experimental,
title = {{Experimental Inter-Report Scatter Bounds Machine-Learning Accuracy for TADF Singlet–Triplet Gaps—Structure and Semi-Empirical Features Reach the Same Limit}},
author = {Jean-Pierre TCHAPET NJAFA and Steve Cabrel Teguia Kouam and Patrick Sorrel Mvoto Kongo and Panebei Samafou and Serge Guy Nana Engo},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010017/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010017/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Experimental Inter-Report Scatter Bounds Machine-Learning Accuracy for TADF Singlet–Triplet Gaps—Structure and Semi-Empirical Features Reach the Same Limit
AU - Jean-Pierre TCHAPET NJAFA
AU - Steve Cabrel Teguia Kouam
AU - Patrick Sorrel Mvoto Kongo
AU - Panebei Samafou
AU - Serge Guy Nana Engo
PY - 2026
DA - 2026/10/07
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010017/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010017/v1
AB - The singlet–triplet gap ∆ESTis the screening parameter for thermally activated delayed fluorescence (TADF), yet independent experimental reports of the same compound scatter by 0.072 eV RMS, and the per-molecule median regressed on is known only to 0.049 eV, about half of any model’s error. This label noise sets a ceiling on every predictor, yet the models sit below it: neither noise nor capacity alone accounts for the remaining error. Against that ceiling, two-dimensional structure predicts as accurately as a semiempirical excited-state calculation: on 231 donor–acceptor molecules (212 Bemis–Murcko scaffolds, scaffold cross-validation), random forests from Morgan fingerprints (MAE 0.091 eV) and RDKit descriptors (0.100 eV) match natural-transition-orbital descriptors from a GFN2-xTB/sTDA excited state (0.096 eV) to within 0.017 eV on paired folds, so the excited-state step buys no accuracy. Two further results are specific to this corpus. First, under a source-paper split the accuracy advantage over a mean-value baseline disappears, but the excited-state descriptors keep the rank correlation the fingerprints lose (paired ∆ρ = 0.19 on a single split, 95 % CI narrowly excluding zero): they order candidates across laboratories without predicting the gap. Second, the leakage that inflates many computed-gap models is learner-dependent—given the S1 and T1 energies as features, a linear model returns ∆EST ≡ ES1 − ET1 exactly (R2 = 1.000) while an axis-aligned forest on the same features recovers only R2 = 0.59 and clears cross-validation unremarked. Progress past the ≈ 0.1eV ceiling needs a curated, solvent-controlled, single-source experimental benchmark—one that would also expose the remaining modelling headroom. Data, models, and code are available (Zenodo DOI: 10.5281/zenodo.17436069).
ER -