Researchers at Los Alamos National Laboratory and the University of New Mexico have developed an AI framework that solves a fundamental physics problem that has vexed scientists for over a century. Calculations that previously required thousands of supercomputer hours now finish in seconds.
The framework, called THOR (Tensors for High-dimensional Object Representation), tackles the configurational integral — a mathematical problem at the heart of understanding how atoms behave inside materials. Published in Physical Review Materials, the results show THOR reproduces the best existing simulations more than 400 times faster.
The Century-Old Problem
The configurational integral captures how particles interact within a material. It’s essential for predicting material properties like phase transitions, thermal behavior, and structural stability. The problem: the integral involves dimensions in the thousands, making direct calculation effectively impossible.
“Classical integration techniques would require computational times exceeding the age of the universe, even with modern computers,” explained Boian Alexandrov, a senior AI scientist at Los Alamos who led the research alongside Duc Truong and UNM professor Dimiter Petsev.
For over a century, scientists have relied on statistical sampling methods — essentially informed guessing — rather than solving the integral directly. These approximations work, but they’re computationally expensive and sometimes unreliable.
How THOR Works
THOR uses a mathematical technique called tensor train cross interpolation to compress enormous high-dimensional problems into manageable pieces. Rather than calculating every possible atomic configuration, the framework identifies patterns and symmetries in crystal structures, dramatically reducing what needs to be computed.
The system integrates machine learning atomic models to predict particle interactions, then exploits the repetitive nature of crystal lattices to simplify calculations further.
Real Results
The research team tested THOR on metals like copper and noble gases under extreme pressure, including crystalline argon. They also calculated tin’s solid-solid phase transition — a notoriously difficult problem in metallurgy.
In each case, THOR matched the accuracy of established Los Alamos simulations while completing the work 400 times faster. Problems that once required dedicated supercomputer time now run on standard research hardware.
“THOR AI opens the door to faster discoveries and a deeper understanding of materials,” said Truong, the lead author.
What This Means
Materials science drives everything from battery technology to aerospace engineering. Faster, more accurate predictions of material behavior could accelerate development of new alloys, superconductors, and structural materials.
The framework is particularly useful for extreme conditions — high pressure, high temperature — where experimental testing is expensive or impossible. Researchers can now screen candidate materials computationally before committing to lab work.
Open Source
THOR is available on GitHub under a BSD-3 license, meaning any research group can use and modify it. The repository includes implementations in Fortran, MATLAB, and Python, making it accessible across different computational environments.
The framework currently handles crystalline solids. Extensions to amorphous materials, liquids, and more complex systems are likely next, though the researchers haven’t announced specific timelines.
The Fine Print
While the speed improvements are dramatic, THOR still requires expertise to use effectively. Setting up tensor network calculations for new materials demands understanding of both the physics and the mathematical framework.
The framework also depends on accurate machine learning models for atomic interactions. For well-studied materials, these exist. For exotic compounds, building the necessary models adds upfront work.
Still, replacing a century of approximations with direct calculations represents genuine progress. When a technique works across copper, argon, and tin, it’s likely to generalize well.