We are excited to announce the publication of a joint paper between the Maginn and Dowling groups:
B. Agbodekhe, M. N. Carlozo, D. O. Abranches, K. D. Jones, A. W. Dowling and E. J. Maginn, “Enhanced Thermophysical Property Prediction with Uncertainty Quantification using Group Contribution-Gaussian Process Regression“, Mol. Syst. Des. Eng., 2025, DOI: 10.1039/D5ME00126A.
Led by Barnabas Agbodekhe, with contributions from Montana Carlozo, Dinis Abranches, and Kyla Jones, this work shows how estimations from simple, first-order group contribution (GC) methods can serve as input features (along with molecular weight) to Gaussian process models for accurate property prediction. This can be used to rapidly estimate physical properties of compounds with much greater accuracy than GC models alone.
Efficient and reliable thermophysical property prediction sits at the heart of any high-throughput computational molecular discovery and design campaign. Thermophysical property predictions from a simple first-order group contribution (GC) model, along with molecular weight (MW), are used as the only two input features to Gaussian process (GP) regression models for enhanced thermophysical property predictions with reliable uncertainty quantification (UQ). Accurate property predictions are obtained with only two input feature dimensions, instead of the tens or hundreds typically used in the literature. The method, known as the GCGP method, provides a state-of-the-art balance of speed, ease of implementation, predictive accuracy, parsimoniousness, and reliable uncertainty quantification. It is especially suited to systems that can be modeled using GC methods, and its scope of applicability can be extended by incorporating other GC methods and/or input features into the GP models. Potential applications of the GCGP method include efficient and enhanced prediction of thermophysical properties with uncertainty quantification for materials discovery via database screening or computer-aided molecular design campaigns.
