Rittig et al. (2026), Molecular machine learning in chemical process design

Fifteen pages, 136 references, four schematic figures and no original data, from the process systems group at RWTH Aachen with EPFL. The subject is end-to-end molecular machine learning: models that learn their own vector representation of a molecule from its structure and map it to a property, as opposed to models fed with hand-picked descriptors.

The comparison of the three model families is the clearest I have read. Graph neural networks pass information along bonds and so assume that properties are set by local atom environments, which the authors liken to group contribution methods made learnable. Transformers attend across all atoms and capture long-range interactions, but they must learn locality from data, which means pretraining and more of it; one molecule has several valid SMILES strings and a transformer may answer differently for each. Matrix completion treats mixture properties as missing entries in a matrix of molecules and cannot say anything about a molecule it has not seen. For mixtures, a learned aggregation of the component vectors lets the mixing effect depend on which molecules are present, which a linear mixing rule cannot do.

The claim that gives the paper its weight is that these models now exceed UNIFAC and COSMO-RS for properties such as binary activity coefficients, when the data suffice. The authors put sufficiency at a few hundred data points, and below that they advise multi-task or transfer learning or a plain descriptor model with XGBoost. In their experience the quantity and quality of data decide accuracy more than the model does. Extrapolation is very limited: an unseen atom type or a temperature outside the training range cannot be predicted reliably, and every prediction is an initial estimate that needs experimental validation.

The section on physics-enriched models sorts the options cleanly. A parallel hybrid predicts the error of a semi-empirical model and keeps its flexibility at the cost of consistency. A sequential hybrid predicts the parameters of the semi-empirical model and inherits its consistency and its ceiling. An embedded hybrid builds the semi-empirical equations into the architecture and trains directly on property data. Physics-informed models add thermodynamic relations to the loss or enforce them through the architecture, with the Helmholtz free energy as the worked example because entropy, internal energy, heat capacity and thermal expansion all follow from its derivatives.

For a process engineer the practical route is the sequential or embedded hybrid: a model that outputs NRTL or PC-SAFT parameters for a molecule with no experimental data, which an existing simulator accepts without any change to the process model. That is the lowest-effort bridge the paper offers between molecular learning and a flowsheet, and the authors are candid that the fuller integration, a learned model inside a process optimization, is still in its infancy because no simulator ships with one.

What to be careful with. No accuracy figures are given for the headline claim; the reader has to follow the citations. The agenda sections on foundation models, benchmarks built with industry and federated learning are proposals. The paper is restricted to string and graph representations and sets 3D geometry aside.

Read for the three-family comparison, the few-hundred-points threshold and the hybrid taxonomy. The last of those is a figure worth redrawing.