Multi-Task Feedforward Neural Networks for Thermodynamic Property Prediction under Small Sample Sizes.

Area of Science:

• Computational Chemistry
• Chemical Engineering
• Materials Science

Background:

• Accurate thermodynamic properties are crucial for reliable explosion modeling of hazardous chemicals.
• Experimental data scarcity necessitates predictive approaches for these properties.
• Existing machine learning quantitative structure-property relationship (ML-QSPR) methods face challenges in multitask accuracy, feature redundancy, and generalization with limited data.

Purpose of the Study:

• To develop a novel machine learning model for accurate thermodynamic property prediction, particularly under small-sample conditions.
• To address limitations of current ML-QSPR methods, including prediction accuracy, feature selection, and generalization.
• To enhance the reliability of explosion modeling through improved thermodynamic property prediction.

Main Methods:

• Proposed a thermodynamically constrained multitask learning network (ThermoMTLnet).
• Leveraged multitask learning to capture correlations among thermodynamic properties.
• Integrated ensemble learning for feature engineering to mitigate overfitting.
• Incorporated physicochemical constraints via a physics-informed neural network (PINN) into the loss function.

Main Results:

• ThermoMTLnet demonstrated superior performance compared to traditional ML models and single-task networks.
• Achieved a 0.23% higher Pearson correlation coefficient (PCC) and 2.44% lower mean absolute error (MAE) on 300-sample datasets compared to the best baseline.
• Maintained consistent superiority on larger datasets, indicating robust generalization capabilities.

Conclusions:

• ThermoMTLnet effectively predicts thermodynamic properties, especially with limited data.
• The model's approach enhances structure-property relationship modeling through multitask learning and physics-informed constraints.
• This work provides a valuable tool for improving explosion modeling and chemical safety assessments.