Thermo Property Models trains two kinds of model for fluids and thermodynamics, and saves either as a model you can predict with.
- Physics-hybrid models start from a group-contribution estimate and learn how far your measurements are from it. A few dozen measured values remove the estimate's bias for your chemistry.
- Mixture models learn a property of a solute in a solvent, or of a single molecule, as a function of temperature and other conditions. One prediction request screens solvents or traces a curve against temperature.
Physics-hybrid models
Pick up to three properties to correct:
| Properties | Estimate |
|---|---|
| Normal boiling point, melting point, critical temperature, pressure and volume, ideal-gas enthalpy and Gibbs energy of formation, enthalpies of vaporization and fusion | Joback group contribution (as in Group-Contribution Properties) |
| Hansen dispersion, polar and hydrogen-bonding parameters | Stefanis–Panayiotou group contribution (as in Hansen Solubility Parameters) |
Give your measured values in whatever unit you use (for example °C or K, kJ/mol or kcal/mol); the estimate is converted to it and the model predicts in it.
For each property, the same panel of learners as QSAR Models is cross-validated on measured minus estimate, and the best one is saved. A prediction is the estimate plus the learned correction, with a 90% prediction interval. When a molecule is unlike anything the correction was trained on, the prediction is flagged; by default the model then returns the estimate alone, unless cross-validation showed the correction still beating the estimate on such molecules. You can also keep the hybrid prediction everywhere.
The training report scores three models on the same held-out molecules: the estimate alone, the hybrid and a pure machine-learning model trained without the estimate. At small dataset sizes the hybrid usually wins clearly; with many measurements the pure model can catch up.
Mixture models
Upload a table with one row per measurement: the solute SMILES (and the solvent SMILES for solutions), the conditions (temperature, and optionally pressure or mole fraction, each with its unit), and one or more measured properties such as log S, an activity coefficient, a density or a viscosity. A message-passing neural network encodes each molecule and appends the conditions, so one model covers every solvent and temperature in your data.
How the test set is drawn decides what the reported error means:
| Split | The test set holds |
|---|---|
| Solute (recommended for new compounds) | Solutes the model never saw |
| Solvent | Solvents the model never saw |
| Both | New solutes in new solvents (rows that mix new and seen molecules are left out) |
| Random | Rows from solutes and solvents also seen in training (the easiest test) |
Predictions take a list of solutes, a list of solvents and condition values, such as a temperature grid in any unit, and return every combination. Each prediction comes with a 90% interval and flags for whether the solute is like the training solutes, the solvent was in training, and the conditions lie inside the training range.
Outputs
| File | Contents |
|---|---|
comparison.csv (physics-hybrid) | The estimate, the hybrid and the pure-ML model on cross-validation and the test set, per property |
test_predictions.csv (physics-hybrid) | Measured values and each model's test-set predictions |
model/ | The saved model with its model card (no pickled code) |
Saved models appear in My Models. Physics-hybrid models can also score whole libraries in a workflow with MPNN Models (Score library).
Related tools
Get the estimates without training from Group-Contribution Properties and Hansen Solubility Parameters; compare a mixture model's solubility curves with the ideal-solution curves of Solid–Liquid Equilibrium / Eutectic Diagram.