JR-MPNN: A Hybrid Thermophysical Property Prediction Model Combining Joback Group Contribution Method and Message Passing Neural Network

In the field of chemical engineering, accurate prediction of thermophysical properties is the foundation of process design. Traditional methods have their own limitations, while JR-MPNN innovatively combines the classic Joback group contribution method (physical prior) with modern message passing neural networks (MPNN, data-driven), retaining interpretability while improving prediction ability, and providing a new solution for thermophysical property prediction.

Thermophysical properties (boiling point, critical temperature, etc.) directly affect chemical process design, but experimental acquisition is costly and time-consuming. Prediction faces three major challenges: complex nonlinear molecular structures, sparse data distribution, and the need to balance accuracy and generalization ability.

The core of the Joback method is that molecular properties are obtained by summing group contributions. Its advantages are simplicity and speed, clear physical meaning, and low data requirements; its limitations are ignoring group interactions and molecular topology, leading to large prediction errors for complex molecules.

MPNN represents molecules as graphs (atoms as nodes, chemical bonds as edges) and captures long-range interactions between atoms through message passing. Its advantage is strong expressive ability; however, pure data-driven models rely on data and have poor interpretability (black box).

JR-MPNN embeds the Joback group contribution as a physical prior into the network (e.g., as an additional input or output constraint), while using MPNN to capture the synergistic effects (hydrogen bonds, conjugated systems, etc.) ignored by the group contribution method, balancing robustness and expressiveness.

Multi-task learning is used to improve data efficiency; the loss function considers dimensional differences and weights, and when data is scarce, Joback predictions are used as soft targets (knowledge distillation); an attention mechanism is introduced to identify key molecular fragments and enhance interpretability.

It can be used for molecular design screening, process simulation data supplementation, and teaching assistance; more importantly, it demonstrates the "physics-inspired machine learning" paradigm, which can be extended to fields such as material properties and reaction kinetics, providing a reference route for researchers.