Autopilot for Digital Materials ✎ express

Original: "AutoPot: Automated and massively parallelized construction of Machine-Learning Potentials"
· Max Hodapp, Guillaume Anciaux
arXiv:2601.01185 · 2026-01-03 · arXiv non-exclusive · independent analysis · 1 min
A new program trains computer models of atoms on its own, saving scientists from the grind.

Computer models that predict material properties at the atomic level usually require long, manual tuning. To work accurately, the model needs to see many examples — how atoms are arranged and how they interact. Scientists spend weeks hand-picking these examples. The new program AutoPot does it all by itself.

AutoPot works like an autopilot: it chooses which atomic configurations to show the model, runs the necessary calculations, and monitors the process. If the model makes a mistake, the program adds new examples to correct the error. AutoPot is built on two tools: BlackDynamite, which performs many calculations simultaneously, and Motoko, which manages their sequence.

The program doesn't just speed up the work: it makes it reproducible. Any scientist can take the same dataset and get the same results.

Thanks to its flexibility, AutoPot is easy to extend: researchers can add their own calculation methods without rewriting the entire code. It's like a construction set where the parts are interchangeable. The program already supports popular modeling methods, and in the future, others can be plugged in.

AutoPot is named after automation and potentials — mathematical functions that describe atomic interactions. So it's 'automatic potential.'

Science fiction often shows computers that design new materials on their own — AutoPot takes the first step toward that by automating the routine part of the work.

Scientists
Erwin SchrödingerHugh Everett IIIAndrey Markov
Concepts
automation machine learned interatomic potential robot Machine Learning parallel processing training learning xgboost Natural Language Processing SLAM quantum algorithm active learning artificial intelligence