(529d) Predictive Parametrization of PC-SAFT EoS By Neural Network Ensembles Based on Molecular Fingerprints

Simulation of chemical or biotechnological processes plays a key role in the design or optimization of industrial processes or production plants. Especially in the design of modern biofuels, physical properties of a variety of different hydrocarbons (ethers, oxygenated hydrocarbons, fatty acid methyl esters) are needed, as the development of new feedstocks for fuel production are a significant challenge of our time. For this purpose, the phase behavior can be efficiently described using advanced equations of state (EoS), such as PC-SAFT. These equations of state (as most thermodynamic models) require a parametrization (pure-component parameters) of each component. This is typically performed by fitting the respective parameters to physical pure-component properties (e.g. vapor pressure, saturated liquid density). However, such experimental data is often not available, especially if complex molecules such as pharmaceuticals are considered. Thus, parametrization using non-experimental inputs are of high interest.
In this work, a neural network (NN)-ensemble was developed to predict PC-SAFT pure-component parameters for non-associating molecules. Beside basic molecular input features (e.g. molar mass, number of rotatable bonds), extended-connectivity fingerprints (ECFPs) were used as key input feature to characterize the molecules. All input features were directly derived from easy available SMILES-notation of the molecules. A dataset comprising ~300 molecules was used for training of the NN-esemble. To increase the statistical validity, 5-fold cross validation with three random initial splits was performed creating a NN-ensemble of 15 NNs to predict the three PC-SAFT pure-component parameters for non-associating molecules. The ensemble prediction yielded good accuracy (AARD < 8 % between ML-predicted PC-SAFT parameters and literature PC-SAFT parameters) for the considered test dataset. Moreover, ML-predicted PC-SAFT pure-component parameters for unknown molecules (not included in training) showed a remarkable performance in describing the experimental (validation) data.
This novel ML-approach offers a reliable and easy access to PC-SAFT pure-component parameters of any non-associating molecule solely based on the chemical structure formulated as SMILES string. Prospectively, ML-predicted pure-component parameters can be used in an early state of process design to estimate phase behavior using a physics-based EoS without additional experimental effort.