Background#
A classical force field is a fixed functional form with parameters fitted to a limited set of reference data, and it fails in characteristic ways on the properties it was never fitted to. Shear viscosity is a good example: standard all-atom parameters systematically underestimate the viscosity of alcohols, because the intermolecular hydrogen bonding that dominates their dynamics is not what the parameters were tuned against. Partition coefficients are another: what a molecule does at a water–octanol interface depends on how its charge distribution responds to the surrounding phase, and a fixed-charge model has no way to express that.
Machine learning is the obvious lever, and the obvious trap. A neural network potential can fit almost anything, which means it can also fit the wrong thing convincingly — reproducing a training set while carrying no transferable physics and offering no way to see where it will break. Our interest is in the narrower and more useful version of the idea: build the physics into the representation, and use learning where a human fitting by hand would be guessing.
One direction is descriptors that encode what the model is missing. Atom-centred symmetry functions describe a local environment through geometry alone; by weighting them with the atomic partial charge and averaging them over a Boltzmann-weighted ensemble of conformers, polarization and conformational entropy enter the representation explicitly rather than being inferred. Trained on measured partition coefficients and summed over atomic contributions, that gives a model that transfers to molecules and datasets it never saw — and one whose predictions can be read back atom by atom.
The other direction runs the opposite way: use learning to interrogate a classical model rather than replace it. Across seventy-five organic liquids spanning nine orders of magnitude in viscosity, feature selection pointed squarely at hydrogen bonding and molecular topology, and showed that the van der Waals term — not the torsional one — was the part carrying the error. Varying that single term systematically was enough to bring the difficult liquids into line. The lesson generalises: knowing which parameter is wrong is worth more than a black box that is right on average.
Both threads feed the same question that runs through our coarse-grained model development — how to build effective interactions that stay predictive away from the conditions they were derived at, and how to know when they stop.
Selected outputs#
- Q. Zhu et al., “Molecular partition coefficient from machine learning with polarization and entropy embedded atom-centered symmetry functions”, Phys. Chem. Chem. Phys. 24 (2022). 10.1039/D2CP02648A
- Q. Zhu et al., “Shear viscosity prediction of alcohols, hydrocarbons, halogenated, carbonyl, nitrogen-containing, and sulfur compounds using the variable force fields”, J. Chem. Phys. 154, 074502 (2021). 10.1063/5.0038267