Accurate formation enthalpies of solids using reaction networks

Crystalline solids play a fundamental role in a host of materials and technologies, ranging from pharmaceuticals to renewable energy. The thermodynamic properties of these solids are crucial determinants of their stability and therefore their behavior. The advent of large density functional theory databases with properties of solids has stimulated research on predictive methods for their thermodynamic properties, especially for the enthalpy of formation Δ_fH. Increasingly sophisticated artificial intelligence and machine learning (ML) models have primarily driven development in this field in recent years. However, these models can suffer from lack of generalizability and poor interpretability. In this work, we explore a different route and develop and evaluate a framework for the application of reaction network (RN) theory to the prediction of Δ_fH of crystalline solids. For an experimental dataset of 1550 compounds we are able to obtain a mean absolute error w.r.t Δ_fH of 29.6 meV atom⁻¹ using the RN approach. This performance is better than existing ML-based predictive methods and close to the experimental uncertainty. Moreover, we show that the RN framework allows for straightforward estimation of the uncertainty of the predictions.