Five Examples — and a Neural Network Predicts Material Properties ✛ ✦
A conventional refrigerator runs on gas that heats the planet. Scientists are looking for a replacement — solid crystals that cool down when compressed. But sifting through thousands of options in the lab is a long process. That's where machine learning comes to the rescue: the neural network, like a seasoned chef, picks up a new recipe from just a few examples. It was shown 5–10 atomic 'snapshots' of ammonium sulfate — a compound of hydrogen, oxygen, and ammonium. It accurately reproduced how atoms rearrange when heated and how entropy (a measure of disorder) changes sharply.
The secret to success lies in hydrogen bonds, invisible bridges that hold atoms together. Accounting for this subtle interaction in computer simulations allowed reliable prediction of the transition temperature using five times fewer computational hours. Surprise: ammonium sulfate — a common nitrogen fertilizer — under pressure becomes a material that cools.
This approach builds on quantum thermodynamics, whose foundations were laid by Max Planck, and the concept of entropy introduced by Ludwig Boltzmann. Fine-tuning an existing neural network opens a fast track to eco-friendly refrigerators without harmful gases.
Just five examples were enough for the neural network to understand when the crystal would change shape — about as many as a child needs to learn a new word.