Physics-Guided Machine Learning for Ionic-Liquid Volumetric Properties | JoVE Visualize

Area of Science:

Physical Chemistry
Computational Chemistry
Materials Science

Background:

Accurate modeling of ionic liquid (IL) volumetric behavior is essential for designing electrolytes used in energy storage and chemical applications.
Classical group contribution methods (GCMs) offer thermodynamic grounding but often lack predictive accuracy and generalizability.
Traditional machine learning (ML) models can struggle with physical consistency and broad applicability.

Purpose of the Study:

To develop a hybrid modeling strategy that integrates classical group contribution methods with physics-informed neural networks (PINN-GCM) for improved IL volumetric property prediction.
To enhance the physical consistency and generalizability of IL modeling by incorporating thermodynamic principles into an ML framework.
To leverage the NIST database for training and validating the novel PINN-GCM model across a wide range of ILs and conditions.

Main Methods:

A hybrid model (PINN-GCM) was developed by coupling a reoptimized Classical-GCM with a PINN.
Thermodynamically optimized parameters from the Tait equation were integrated into the PINN's hybrid loss function.
The model was trained and tested on experimental data for 92 ILs, covering temperatures from 217-473 K and pressures from 0.1-207 MPa.

Main Results:

The PINN-GCM framework achieved high accuracy, with aggregate RAAD values of 0.067% (training) and 0.065% (test) at the IL level.
Ion-level models demonstrated strong combinatorial generalization to new IL pairings, with extrapolation on 2,958 data points from 21 unseen combinations.
The model's applicability is validated within the experimental temperature-pressure range and requires ions present in the established library.

Conclusions:

The hybrid PINN-GCM strategy effectively models the volumetric behavior of ionic liquids, merging physics-based principles with machine learning.
This approach shows significant promise for improving electrolyte design in energy storage and other chemical systems.
Future work may involve extending the framework to predict other IL properties like viscosity and conductivity and integrating it into process simulation tools.