Los Alamos AI Framework Solves a 100-Year-Old Physics Problem in Seconds Instead of Weeks

A new artificial intelligence framework developed by scientists at Los Alamos National Laboratory and the University of New Mexico has solved one of the most intractable computational problems in materials science — one that has resisted solution for roughly a century — in seconds rather than the weeks of supercomputer time previously required. The framework, called THOR, targets what specialists call the "configurational integral" problem: the set of calculations needed to predict how atoms in a material will arrange themselves and how that arrangement determines the material's thermodynamic and mechanical properties.

The configurational integral matters enormously in practical science. When engineers want to design new alloys that can withstand extreme heat, predict how a material will behave under the crushing pressures found inside a nuclear reactor, or understand the phase transitions that govern how a substance melts or crystallizes, they need to evaluate the configurational integral. The problem is that the number of possible atomic configurations in even a small sample of material is astronomically large — the calculation scales exponentially with the number of atoms, making brute-force computation impractical.

THOR attacks this problem with a combination of three advanced techniques. First, it uses tensor network algorithms — specifically an approach called "tensor train cross interpolation" — to represent the enormous configuration space in a compressed form that discards redundant information. Second, it incorporates machine learning potentials, which are neural networks trained to predict the energy of any given atomic configuration far faster than conventional quantum mechanical calculations. Third, it exploits the crystalline symmetry of materials to reduce the number of configurations that need to be evaluated by orders of magnitude. The combination of these three techniques achieves a speedup of more than 400 times over traditional methods.

The team validated THOR on three different test cases: copper, crystalline argon subjected to extreme pressures, and tin undergoing phase transitions. In each case, THOR's predictions matched those of established but slow methods while completing the calculation in seconds on a standard computing cluster. Lead researcher Duc Truong of Los Alamos National Laboratory said the result opens the door to material design workflows that have been effectively impossible until now. "Extreme-pressure phases, complex alloys, and materials that undergo transitions under operating conditions have all been out of reach," Truong said. "THOR changes that."

Senior AI scientist Boian Alexandrov, who led the broader research program at Los Alamos, said the configurational integral problem's century-long resistance to solution was not for lack of effort. "The problem has been considered computationally intractable at scale since the 1920s," he said. "Every decade, people have tried new approaches. What's different now is the convergence of machine learning potentials, tensor algebra, and crystal symmetry exploitation — no single one of those was sufficient, but together they crack the problem."

The results were published in Physical Review Materials in March 2026, and the team has released THOR's code as open source on GitHub. Alexandrov said potential applications include designing heat-resistant materials for jet turbines and fusion reactors, identifying new battery cathode materials, and accelerating drug development by predicting how pharmaceutical compounds will crystallize. The Los Alamos team is now collaborating with industrial partners to deploy THOR on real-world design challenges.