COSMO-NET: fast and accurate machine learning surrogates for COSMO-based molecular descriptors.

Accurate predictive thermodynamic models are essential tools for the computational design of new molecules. Models based on COnductor-like Screening Models (COSMO), such as COSMO-SAC and COSMO-RS, are well-suited for this purpose but require computationally expensive quantum-mechanical (QM) calculations to generate key properties - the surface charge density distribution (p(σ)), the surface area (A) and the cavity volume (V) - for each molecule. Here we develop graph-based neural network surrogate models, based on a directed message passing neural network (DMPNN) and a graph convolutional network (GCN), to predict these molecular properties rapidly and accurately. We train the models on a dataset of over 16 000 compounds generated with an automated QM calculation pipeline. The DMPNN model outperforms GCN for the prediction of the surface charge density, while the GCN provides a higher accuracy for surface area and cavity volume. We build on these strengths to propose a hybrid model: COSMO-NET. When applied to predict octanol-water partition coefficients and solid-liquid solubility across diverse solvents, COSMO-NET consistently outperforms individual models with the ranking: COSMO-NET > DMPNN > GCN > ECFP-MLP. COSMO-NET demonstrates remarkable robustness with minimal performance variation across different data splitting strategies. These results demonstrate that machine-learning surrogates can reliably replace costly QM calculations while maintaining accuracy, enabling the rapid discovery of new molecules and accurate evaluation of their thermodynamic properties.