THOR AI just compressed a 100-year materials simulation from weeks to seconds — a 400x speedup over Los Alamos supercomputer runs.

Released March 15 by researchers at the University of New Mexico and Los Alamos National Laboratory, the Tensors for High-dimensional Object Representation framework rewrites how the field computes configurational integrals. Those integrals describe how particles interact inside metals, ceramics, and gases under extreme conditions. Direct solution had been called impossible since the 1920s.

Los Alamos senior AI scientist Boian Alexandrov led the project.

"The configurational integral — which captures particle interactions — is notoriously difficult and time-consuming to evaluate, particularly in materials science applications involving extreme pressures or phase transitions," Alexandrov said. "Accurately determining the thermodynamic behavior deepens our scientific understanding of statistical mechanics and informs key areas such as metallurgy."

Why is the integral so hard?

The obstacle has a name — the "curse of dimensionality." The integrand spans thousands of variables; each added variable multiplies the work. Classical integration on a problem this large would need longer than the age of the universe, even on today's biggest machines. UNM chemical and biological engineering professor Dimiter Petsev, a frequent Alexandrov collaborator, called direct solution "considered impossible" until tensor network methods emerged as a benchmark.

For decades, materials researchers leaned on indirect sampling — molecular dynamics and Monte Carlo simulations — to estimate particle behavior. THOR skips the sampling. It solves the integral directly.

How does THOR do it?

The framework wraps the integrand into a "tensor train" — a chain of small, connected matrices — using an algorithm called tensor train cross interpolation. A specialized version detects crystal symmetries in the material and skips redundant computation.

In published benchmarks on three test systems:

• Copper metal at standard conditions
• Crystalline argon under extreme pressure
• The solid-solid phase transition in tin

THOR reproduced prior Los Alamos simulation results while running more than 400 times faster. The team published the work in Physical Review Materials, 2025; 9 (8), DOI: 10.1103/xrbw-xr49.

Lead author Duc Truong framed the result as a reset.

"This breakthrough replaces century-old simulations and approximations of configurational integral with a first-principles calculation," Truong said. "THOR AI opens the door to faster discoveries and a deeper understanding of materials."

THOR also pairs with modern machine learning atomic potentials, so researchers can swap in different chemistry without rewriting the pipeline. The framework is available on GitHub for outside testing.

What hasn't been tested yet?

The paper covers three crystalline systems. It skips amorphous materials and high-entropy alloys — the disordered regimes that have stumped metallurgists for decades. Whether tensor compression survives those workloads will decide if THOR becomes a foundry tool or stays a research curiosity.

For now, the math is finally tractable. Chipmakers searching for new interconnect metals, battery designers chasing denser cathodes, and aerospace alloy houses all face the same integral. The same 400x head start is now theirs to verify — or refute — before the summer materials conferences arrive.