Skip to content
DraconiticResearch
Search Sources API
201,818 studies31 sourcesUpdated 13 min ago
PreprintCC BYChemRxiv

The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions

  • Alexander Kanzow1,2
  • Cesare Boriosi1,2
  • Carlos R. Jacinto-Mejía1,2
  • Loriano Storchi3,4,5
  • Giovanni Bistoni1,2
5 affiliations
  1. 1Department of Chemistry, Biology and Biotechnology
  2. 2University of Perugia
  3. 3Dipartimento di Farmacia
  4. 4Università G. D’Annunzio
  5. 5Institute of Science and Chemical Technologies "Giulio Natta"-SCITEC, National Research Council, via Elce di Sotto, Perugia, Italy, 06123
Open at sourceDOICite

Abstract

Machine-learned formation enthalpies for inorganic materials approach density functional theory (DFT) accuracy, yet thermodynamic stability predictions derived from them can remain unreliable, a gap commonly associated with error cancellation in DFT that independent per-compound training does not reproduce. We introduce the interference index ξ, a dimensionless measure of how strongly errors reinforce or cancel in linearly derived quantities. Its root-mean-square average, ξrms, provides a reaction-size-independent null reference, with ξrms = 1 for the isotropic null model. Across seven published models benchmarked on Materials Project compounds, error-cancellation behaviour is distinct from per-compound accuracy and becomes progressively less favourable with increasing formation-enthalpy accuracy among the compositional models. We then introduce iiLoss, a reaction-aware objective that favours cancellation of formation-enthalpy errors rather than directly minimising decomposition-enthalpy error. On Materials Project data, iiLoss reduces decomposition-enthalpy error by 21% and improves thermodynamic stability classification by 6.8% in F1 score, outperforming direct supervision on decomposition enthalpies. These results show that error cancellation can be quantified and directly optimised to improve derived thermodynamic predictions.

Related

  1. Model Dependence and Cancellation in a Fixed-Geometry Lactone Energy Cycle
    Situo Xing
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  2. Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
    Emilio Nuñez-Andrade, Isaac Vidal-Daza, Rafael Gomez-Bombarelli +2
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  3. Statistical predictions on the encapsulation of single molecule binding pairs into sized-dispersed nanocontainers
    Guillaume Longatte, Fabio Lisi, Xueqian Chen +6
    Physical Chemistry Chemical PhysicsPeer-reviewedLicence not statedFull text1 citedAnalytical chemistryHAL
  4. Crystal structure, Hirshfeld surface analysis, computational study, molecular docking predictions and biological evaluation of a new Schiff base (E)-2-((2,4-dichlorobenzylidene)amino)aniline
    Raounek Rouag, Amani Hind Benahsene, Mohamed Abdesselem Dems +4
    Journal of Molecular StructurePeer-reviewedLicence not stated2 citedOrganic chemistryHAL
  5. Universal Model for Atoms (UMA) in Computational Asymmetric Catalysis: Comparing Enantioselectivity Predictions
    Soumik Das, Cyndi Qixin He, Yu-hong Lam
    ChemRxivPreprintCC BYFull textOrganic chemistryChemRxiv

Cite

BibTeXRIS
BibTeX
@misc{kanzow2026interference,
  title = {{The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions}},
  author = {Alexander Kanzow and Cesare Boriosi and Carlos R. Jacinto-Mejía and Loriano Storchi and Giovanni Bistoni},
  howpublished = {ChemRxiv},
  year = {2026},
  month = sep,
  doi = {10.26434/chemrxiv.15008797/v1},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008797/v1},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions
AU  - Alexander Kanzow
AU  - Cesare Boriosi
AU  - Carlos R. Jacinto-Mejía
AU  - Loriano Storchi
AU  - Giovanni Bistoni
PY  - 2026
DA  - 2026/09/14
PB  - ChemRxiv
DO  - 10.26434/chemrxiv.15008797/v1
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008797/v1
AB  - Machine-learned formation enthalpies for inorganic materials approach density functional theory (DFT) accuracy, yet thermodynamic stability predictions derived from them can remain unreliable, a gap commonly associated with error cancellation in DFT that independent per-compound training does not reproduce. We introduce the interference index ξ, a dimensionless measure of how strongly errors reinforce or cancel in linearly derived quantities. Its root-mean-square average, ξrms, provides a reaction-size-independent null reference, with ξrms = 1 for the isotropic null model. Across seven published models benchmarked on Materials Project compounds, error-cancellation behaviour is distinct from per-compound accuracy and becomes progressively less favourable with increasing formation-enthalpy accuracy among the compositional models. We then introduce iiLoss, a reaction-aware objective that favours cancellation of formation-enthalpy errors rather than directly minimising decomposition-enthalpy error. On Materials Project data, iiLoss reduces decomposition-enthalpy error by 21% and improves thermodynamic stability classification by 6.8% in F1 score, outperforming direct supervision on decomposition enthalpies. These results show that error cancellation can be quantified and directly optimised to improve derived thermodynamic predictions.
ER  -