An Adaptive ML Framework to Predict PC-SAFT Parameters for Mixtures

The application of perturbed-chain statistical associating
fluid
theory (PC-SAFT) to complex mixtures requires fitting pure-component
parameters and determining binary interaction parameters for all molecular
pairs. Since these binary parameters depend on the pure-component
sets, an integrated machine learning (ML) framework was developed
to predict them directly from pure-component inputs. This allows flexible
variation of component parameters while maintaining consistency. A
neural network ensemble was trained on 7300 binary interaction parameters
from 6301 systems, achieving an average mean absolute error (MAE)
of 0.0096 for kij predictions.
Using ML-predicted kij values, vapor–liquid (VLE) and liquid–liquid equilibria
(LLE) for binary and ternary systems were successfully reproduced.
The results demonstrate that the proposed ML framework enables efficient
and accurate PC-SAFT predictions with minimal experimental input,
providing a powerful tool for early stage process development, where
reliable thermodynamic modeling is needed without extensive data acquisition.