AI solved a century-old physics problem in a matter of seconds

Researchers from the University of New Mexico and Los Alamos National Laboratory have developed a new computational method that could solve one of the most challenging problems in statistical physics — calculating so-called configurational integrals. The system has been named THOR AI (Tensors for High-dimensional Object Representation). The work has been published in the journal Physical Review Materials (PRM).

Configurational integrals are used to describe particle interactions and allow predicting thermodynamic and mechanical properties of materials. Such calculations are necessary, for example, for studying phase transitions, the behavior of substances under extreme pressure, and the development of new materials. However, computing these integrals directly is extremely difficult.

"The configurational integral, which describes particle interactions, is extraordinarily difficult to compute, especially in materials science problems involving high pressures or phase transitions," explained the project leader, an artificial intelligence specialist at Los Alamos National Laboratory, Boian Alexandrov.

The main problem lies in the so-called "curse of dimensionality." As the number of variables increases, computational complexity grows exponentially. Even modern supercomputers struggle with such tasks. For this reason, scientists have spent decades using approximate methods — such as molecular dynamics or the Monte Carlo method — which simulate atomic motion and require enormous computational resources.

The new approach enables such calculations to be performed directly. The THOR AI algorithm uses tensor network methods — a mathematical technique that allows representing vast multidimensional data as a set of interconnected simpler elements.

The THOR AI system breaks down a complex problem into a sequence of more compact computations and applies a method known as tensor interpolation. Additionally, the algorithm is capable of automatically identifying symmetries in the crystal structure of materials, which further reduces the volume of computations.

As a result, calculations that previously took thousands of hours can now be performed in seconds without loss of accuracy.

The researchers tested the new system on several materials, including copper, crystalline argon under high pressure, and complex phase transitions of tin. In all cases, the results matched data from more labor-intensive simulations, yet the computations were performed more than 400 times faster.

Furthermore, THOR AI can be integrated with modern machine learning models that describe atomic interactions. This makes it possible to analyze material behavior under a wide variety of physical conditions.

According to the researchers, the new technology could accelerate materials development and deepen understanding of fundamental processes in physics, chemistry, and materials science.