Solid-Phase Standard Enthalpies of Formation from SMILES with a Calibrated Per-Molecule Uncertainty: Application to a 13 606-Compound CHONFCl Inventory
- 1R&D Mediation, 71 avenue de Lattre de Tassigny, Bourges, France, F-18000
Abstract
The standard enthalpy of formation of a solid, ∆fH ◦(s, 298 K), governs the detonation performance and the thermal hazard of an energetic material, yet it is measured for only a few hundred compounds, and 30 % of the compounds measured more than once carry determinations that disagree by more than 20 kJ mol−1. I report a route from a SMILES string to ∆fH ◦(s, 298 K) that returns, together with the value, a molecule-specific standard deviation calibrated against experiment. The gas-phase enthalpy is obtained from the ANI-2x neural network potential evaluated on a connectivity-based hierarchy (CBH-2, with CBH-1 fallback) isodesmic balance in which the nitro group is treated as a single pseudo-atom, without which the balance cannot be closed on a charge-separated nitrogen; a rigid-rotor harmonic-oscillator correction supplies the thermal terms. The sublimation enthalpy is a 32-descriptor additive model fitted to 229 measurements harvested from the NIST Chemistry WebBook. Two systematic shifts that survive leave-one-out validation are removed. The uncertainty is a heteroscedastic Gaussian variance model. The whole chain is validated by a nested leave-one-out procedure in which the sublimation model, the acceptance threshold, the systematic shifts and the variance model are all re-estimated without the compound they predict, and that compound’s own gas-phase reference is withheld: over 281 compounds the mean absolute error is 26.1 kJ mol−1, the mean |z| is 0.80 against 0.80 expected for a well-calibrated Gaussian, and the coverage is 71.2 % at one standard deviation and 93.6 % at two, against 68.3 % and 95.4 % expected. For scale, independent experimental determinations of the same solid deviate from their own median by 7.6 kJ mol−1 on average, so two determinations of one compound typically differ by about twice that. The procedure was applied to an inventory of 13 606 CHONFCl compounds; 13 267 (97.5 %) were predicted in 314 core-hours on a single desktop machine, the remainder being open-shell radicals outside the domain of ANI-2x. I state explicitly, and quantify, the size beyond which the uncertainty model itself extrapolates.
Related
- Selenenic vs. sulfenic acids: Enthalpies of formation, bond dissociation energies, and thermochemical properties from composite methods
- Hydride formation and interfacial enthalpies between hydrogen and Mg mixtures with immiscible elements/alloys
- 1H NMR-Based Framework for Estimating Molecular Formation Enthalpies of Organic Compounds.
- Physics-Guided Learning with Calibrated Uncertainty for Molecular Thermodynamics Beyond Observed Chemical Space
- Seed layer formation by deposition of micro-crystallites on a revolving substrate. Part II: application to some nonlinear effects
BibTeX
@misc{brunet2026solidphase,
title = {{Solid-Phase Standard Enthalpies of Formation from SMILES with a Calibrated Per-Molecule Uncertainty: Application to a 13 606-Compound CHONFCl Inventory}},
author = {Luc E. Brunet},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009244/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009244/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Solid-Phase Standard Enthalpies of Formation from SMILES with a Calibrated Per-Molecule Uncertainty: Application to a 13 606-Compound CHONFCl Inventory
AU - Luc E. Brunet
PY - 2026
DA - 2026/09/22
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009244/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009244/v1
AB - The standard enthalpy of formation of a solid, ∆fH ◦(s, 298 K), governs the detonation performance and the thermal hazard of an energetic material, yet it is measured for only a few hundred compounds, and 30 % of the compounds measured more than once carry determinations that disagree by more than 20 kJ mol−1. I report a route from a SMILES string to ∆fH ◦(s, 298 K) that returns, together with the value, a molecule-specific standard deviation calibrated against experiment. The gas-phase enthalpy is obtained from the ANI-2x neural network potential evaluated on a connectivity-based hierarchy (CBH-2, with CBH-1 fallback) isodesmic balance in which the nitro group is treated as a single pseudo-atom, without which the balance cannot be closed on a charge-separated nitrogen; a rigid-rotor harmonic-oscillator correction supplies the thermal terms. The sublimation enthalpy is a 32-descriptor additive model fitted to 229 measurements harvested from the NIST Chemistry WebBook. Two systematic shifts that survive leave-one-out validation are removed. The uncertainty is a heteroscedastic Gaussian variance model. The whole chain is validated by a nested leave-one-out procedure in which the sublimation model, the acceptance threshold, the systematic shifts and the variance model are all re-estimated without the compound they predict, and that compound’s own gas-phase reference is withheld: over 281 compounds the mean absolute error is 26.1 kJ mol−1, the mean |z| is 0.80 against 0.80 expected for a well-calibrated Gaussian, and the coverage is 71.2 % at one standard deviation and 93.6 % at two, against 68.3 % and 95.4 % expected. For scale, independent experimental determinations of the same solid deviate from their own median by 7.6 kJ mol−1 on average, so two determinations of one compound typically differ by about twice that. The procedure was applied to an inventory of 13 606 CHONFCl compounds; 13 267 (97.5 %) were predicted in 314 core-hours on a single desktop machine, the remainder being open-shell radicals outside the domain of ANI-2x. I state explicitly, and quantify, the size beyond which the uncertainty model itself extrapolates.
ER -