PINN-SAC consists of two core components: a sigma profile predictor and a segment activity coefficient predictor. The first stage utilizes a pre-trained SMILES encoder to predict sigma profiles with high accuracy (R² ≈ 0.97) based on a dataset of approximately 25,000 molecular profiles. In the second stage, over a million theoretically computed segment activity coefficients are used to train physics-informed neural networks (PINNs), which enforce thermodynamic consistency by satisfying the Gibbs-Duhem equation and symmetry constraints. Finally, the model undergoes fine-tuning with targeted activity coefficient data, enhancing its predictive accuracy for specific systems.
The base PINN-SAC model achieved an R² of 0.92 on binary activity coefficients on the VT-2005 dataset, comparable to COSMO-SAC results. PINN-SAC offers substantial advantages in efficiency and usability, eliminating the need for quantum mechanical and iterative statistical calculations. Furthermore, with sufficient experimental data, PINN-SAC can be fine-tuned to surpass COSMO-SAC in specific datasets. Its fast inference, thermodynamic rigor, and flexibility make it well-suited for process simulations, industrial design, and high-throughput screening applications, allowing adaptation to specific datasets and outperforming conventional thermodynamic models.