A translator who has memorized all the lines in advance responds instantly — no need to listen to the whole conversation again. That's how the new model works: instead of running atoms through a computer for hours, it immediately outputs the crystal's free energy at any temperature and pressure.

Previously, to find out which phase of a material is stable when heated, you had to run molecular dynamics — a computer simulation of thousands of atoms in motion. A single such calculation could take hours on a supercomputer. And for each new temperature, the whole process started from scratch.

The authors took a different route. They took a universal machine-learned potential that could already compute the energy of a crystal at zero temperature and added a lightweight extension. This extension learns from data at three levels of accuracy: first, a cheap approximation where atoms vibrate like independent springs; then, accurate simulations; and finally, experimental measurements of heat capacity and equations of state.

The result: from a single static structure, the model outputs the entire Gibbs free energy surface as a function of temperature and pressure. And along with it, the derivatives: volume, entropy, heat capacity, and elastic modulus. It's as if the translator not only knew the answers but could also predict how the conversation would turn with any change of topic.

The model was tested on 255 compounds from the NIST–JANAF reference data. The heat capacity error was 1.94 J/(K·mol atoms)—at the level of experimental accuracy. It predicts polymorphic transitions in titanium, hafnium, and CaSiO3, including phases that are unstable at zero temperature. And after calibration on experimental data, it corrected the systematic overestimation of volume inherited from the base method and recovered the correct phase diagram of SiO2.

Now the translator who has learned all the lines is ready to work without stumbling — even where you used to have to wait for hours.