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

Reliable explosion modeling requires accurate thermodynamic properties of hazardous chemicals, yet the scarcity of experimental data highlights the urgent need for predictive approaches to bridge this critical gap. In recent years, machine learning-assisted quantitative structure-property relationship (ML-QSPR) approaches have been widely adopted for thermodynamic property prediction due to their high computational efficiency and strong interpretability. Nevertheless, current methods still suffer from limited multitask prediction accuracy, redundant features, and poor generalization under small-sample conditions and descriptor calculation constraints. In this work, we propose a thermodynamically constrained multitask learning network designed for small-sample thermodynamic property prediction. The model leverages multitask learning to exploit inherent correlations among thermodynamic properties, integrates ensemble learning strategies into the feature engineering process to alleviate overfitting, and incorporates physicochemical constraints into the loss function by constructing a physics-informed neural network (PINN) to enhance structure-property relationship modeling. Comparative experiments demonstrate that ThermoMTLnet outperforms traditional machine learning models and single-task neural networks across multiple metrics, including Pearson correlation coefficient (PCC) and mean absolute error (MAE). Furthermore, ThermoMTLnet achieved a 0.23% higher PCC and 2.44% lower MAE than the best baseline on 300-sample data sets, while maintaining consistent superiority on larger data sets. This demonstrates its robust generalization capability for thermodynamic property prediction with limited samples.