Universal machine-learning interatomic potentials are trained on massive datasets spanning hundreds of elements and thousands of compounds. They're supposed to work everywhere — any combination of atoms, any crystal structure, any composition. The promise is that you train once and deploy universally.
Zhou and Komsa tested these universal models on two-dimensional high-entropy alloy sulfides — specifically (Mo,Ta,Nb,W,V)S₂, a system that has been experimentally synthesized. Every universal model they tested produced inadequate predictions for mixing energy. Not slightly off. Qualitatively wrong — unable to correctly rank which compositions are stable and which aren't.
The fix is fine-tuning: taking the universal model and retraining it on a curated set of configurations specific to the target chemistry. After fine-tuning, the models achieve near-DFT accuracy. They can then run Monte Carlo simulations and explore random structures at scales that density functional theory cannot reach. The universal model provides the foundation; the fine-tuning provides the accuracy.
The structural lesson is about what “universal” means in machine learning. It doesn't mean “works everywhere out of the box.” It means “contains enough general knowledge to be efficiently specialized.” The universal model isn't a finished tool — it's a starting material. Like a blank key that has the right profile but needs to be cut for the specific lock.
This gap between “universal” and “actually works” is underappreciated. The marketing is universality. The practice is fine-tuning. The value of the universal model is real — training from scratch for every new chemistry would be prohibitively expensive. But the value is in reducing the cost of specialization, not in eliminating it. The general model makes the specific model cheaper. It doesn't make it unnecessary.