Effect of the Clapeyron loss on prediction accuracy

Determine whether the Clapeyron loss in the soft-constrained Clapeyron-GNN improves prediction accuracy, rather than only improving the thermodynamic consistency of predictions, when training data are unevenly available across the four vapor–liquid-equilibrium properties.

Background

The hard-constrained Clapeyron-GNN calculates the enthalpy of vaporization from the other predicted properties, whereas the soft-constrained model includes the Clapeyron relation as a loss term. Because the Clapeyron loss is computed from the predictions, it provides a loss signal for all four properties even when experimental data are missing for some of them. The paper explicitly leaves unresolved whether this additional loss signal improves predictive accuracy or merely enhances adherence to the thermodynamic relation.

References

However, it is unclear if this also helps with prediction accuracy or only affects consistency of predictions.

— A systematic comparison of hard- and soft-constrained physics-informed molecular machine learning with the Clapeyron equation  (2609.36947 - Pavšek et al., 29 Sep 2026) in Section 3.2, subsection “Soft-constrained vs. hard-constrained Clapeyron-GNN”

As the RMSE of the hard-constrained model is lower for the vapor pressure but higher for the molar volumes than the RMSE of the soft-constrained one for these two molecules, it is not clear which model follows the true trend more closely in this region.

— A systematic comparison of hard- and soft-constrained physics-informed molecular machine learning with the Clapeyron equation  (2609.36947 - Pavšek et al., 29 Sep 2026) in Section 3.3, subsection “Prediction for individual molecules”