The neural network tuner: how five quantum calculations open the door to Freon-free fridges ✛ ✦
Today's Freon-based refrigerators and air conditioners silently poison the atmosphere. An alternative is solid-state caloric materials that cool under rapid pressure changes via the barocaloric effect. At its core lies a sharp jump in entropy during a phase transition — a concept introduced by Ludwig Boltzmann and inextricably linked to the second law of thermodynamics. To find such substances among thousands of candidates, precise atomistic simulations are needed, but each quantum-mechanical DFT calculation costs as much as manufacturing a new grand piano.
Training machine-learned interatomic potentials (MLIP) usually requires thousands of DFT snapshots, like building a piano from scratch. But the authors took the path of fine-tuning: like an experienced tuner who doesn't rebuild the instrument, but merely tightens a few strings, they started with the pre-trained model MACE-MPA-0, which had absorbed the quantum harmony of many molecules. By fine-tuning it on just 5–10 configurations of ammonium sulfate, a crystal where ammonium ions (NH4) and sulfate weave a network of hydrogen bonds involving hydrogen and oxygen, they reproduced the structural phase transition over the range 100–400 K.
Training from scratch required at least 50 samples to notice a jump in lattice parameters — otherwise the transition sounded off-key. Fine-tuning, drawing on the rich 'musical memory' of the base model, captured the transformation moment with tuning-fork precision. A curious nuance emerged: without dispersion corrections accounting for van der Waals interactions, the transition was not reproduced in any of the models. This is like a forgotten chord: hydrogen bonds rearrange the entire 'harmony' of the crystal, and it is this process that underlies the barocaloric effect. Ironically, the same hydrogen bonds that hold together the double helix of DNA here act as molecular springs for cooling. Here the quantum precision of methods dating back to Max Planck intertwines with the subtle interplay of quantum thermodynamics and the real macroworld.
This extreme data economy is not just a record. The freed-up computing power can be poured into more accurate quantum methods, like hybrid functionals, or into molecular dynamics of millions of atoms. Like a tuner who, having quickly handled a concert grand piano, tackles fine voicing, materials scientists gain a chance to sift through thousands of candidates in search of perfect solid-state refrigerants. On the horizon — extending the approach to magnetocaloric and electrocaloric systems, where magnetic and electric 'instruments' will join the mechanical 'tuning,' and integration with experiment for direct property prediction.
The original model would have misjudged the transition temperature by tens of degrees, but five fine-tuning examples reduced the error to 60 K (compared to the experimental value of 219 K)—like a string tuned to within a quarter tone.