Balancing Molecular Information and Empirical Data in the Prediction of Physico-Chemical Properties

• DOI: 10.48550/arxiv.2406.08075
• Publication Date: 2024-06-12
• Authors: Johannes Zenn; Dominik Gond; Fabian Jirasek; Robert Bamler
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG); 004

Predicting the physico-chemical properties of pure substances and mixtures is a central task in thermodynamics. Established prediction methods range from fully physics-based ab-initio calculations, which are only feasible for very simple systems, over descriptor-based that use some information on molecules to be modeled together with fitted model parameters (e.g., quantitative-structure-property relationship or classical group contribution methods), representation-learning methods, may, extreme cases, completely ignore molecular descriptors extrapolate existing data property matrix completion methods). In this work, we propose general method combining representation learning using so-called expectation maximization algorithm probabilistic machine literature, uses uncertainty estimates trade off between two approaches. The proposed hybrid exploits chemical structure graph neural networks, but it automatically detects cases where structure-based predictions unreliable, case corrects them by based can better specialize unusual cases. effectiveness demonstrated activity coefficients binary as an example. results compelling, significantly improves predictive accuracy current state art, showcasing its potential advance general.