Configurational Integrals Solved by Tensorization [Slides]

Boian S. Alexandrov, Duc P. Truong, Derek DeSantis, Benjamin Nebgen2023 · DOI

Introduction

In the field of computational materials science, determining the thermodynamic properties of a substance—such as entropy, free energy, or pressure—requires solving the configurational integral ($Z$). For a system containing $N$ atoms, this is a $3N$-dimensional integral that accounts for all possible spatial arrangements weighted by their Boltzmann factor ($e^{-eta U}$).

As $N$ grows, the "curse of dimensionality" makes standard numerical integration impossible. For a modest system of 500 atoms (1,500 dimensions), a standard grid of 400 points per dimension would require $400^{1500}$ operations—a number that vastly exceeds the number of atoms in the observable universe. Consequently, researchers typically rely on stochastic methods like Molecular Dynamics (MD), which can be computationally expensive and often struggle to calculate absolute thermodynamic quantities like entropy.

Methodology: The Tensorization Approach

Researchers at Los Alamos National Laboratory (LANL) have developed a breakthrough method that utilizes Tensorization to break these high-dimensional problems into a series of simple, one-dimensional calculations. The approach relies on Tensor Train (TT) cross-interpolation and the Maximum Volume (Maxvol) principle.

Rank-1 Decomposition and Coordinate Transformation

The core innovation lies in approximating the $3N$-dimensional integrand as a low-rank tensor. Specifically, the method seeks a rank-1 decomposition where the high-dimensional function is treated as a product of $3N$ univariate functions:

$$f(x_1, \dots, z_N) \approx f_1(x_1) \cdot f_2(y_1) \cdot \dots \cdot f_{3N}(z_N)$$

To ensure this approximation is physically accurate, the authors apply a coordinate transformation. By shifting the system so that equilibrium atomic positions are at the origin and the minimum potential energy is zero, the high-dimensional energy landscape becomes highly separable. The Maxvol principle is then used to identify the most significant regions of the function to optimize the decomposition.

Exploiting Symmetry

In crystalline materials, many atoms exist in identical environments. The researchers found that the 1D integrals for these atoms are often identical. In perfect lattices, this symmetry allows a 1,500-dimensional problem to be reduced to just a few unique 1D integrals, drastically cutting computational costs.

Validation and Results

The method was validated using several physical systems, ranging from simple noble gases to complex metals.

  • Solid Argon: Using a Lennard-Jones potential for a 500-atom system, the method reduced the required grid points from the theoretical $400^{1500}$ to just 400. The resulting energy and pressure calculations showed near-perfect agreement with traditional MD simulations across a temperature range of 50K to 300K.
  • Liquid Argon: The approach was tested on disordered systems to evaluate the limits of the rank-1 approximation. Even without lattice symmetry, the method remained effective by computing individual 1D integrals for each particle.
  • Tin (Sn): To demonstrate versatility, the team applied the method to Tin using complex Machine Learned Interatomic Potentials (MLIAP). The algorithm successfully approximated energies across four different crystal structures (FCC, BCC, BCT, and Diamond Cubic), proving it can handle sophisticated modern potentials.

Conclusion and Future Outlook

By transforming exponential complexity into linear complexity, tensorization provides a powerful new tool for materials modeling. While the current rank-1 approximation is highly effective for solids, future work will explore rank-2 decompositions to handle multi-component systems like Sodium Chloride (NaCl) and capture more complex many-body correlations.

The ultimate goal is to apply this technique to construct full phase diagrams and simulate dynamic processes, such as the charge/discharge cycles of lithium-ion batteries, by rapidly calculating free energy changes that were previously computationally out of reach.

Summary updated

Abstract

e x i t > e U (x 1 ,...,z N ) dx 1 . .
BibTeX
@misc{alexandrov2023configurational,
  author = {Boian S. Alexandrov and Duc P. Truong and Derek DeSantis and Benjamin Nebgen},
  title = {{Configurational Integrals Solved by Tensorization [Slides]}},
  year = {2023},
  doi = {10.2172/2350587},
}
Figure