The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions
Year: 2026
Type: preprint
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.
Source: ChemRxiv
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213231093