Time- and memory-efficient representation of complex mesoscale potentials

Grigoriy M. Drozdov, Igor A. Ostanin, Ivan V. Oseledets

Abstract

Mesoscale modeling often requires calculating high-dimensional potential energy surfaces for complex-shaped nanostructures, leading to a "curse of dimensionality" in memory storage. This paper introduces a method using Tensor Train (TT) decomposition to represent these potentials as a product of low-rank matrices. Using carbon nanotubes as a test case, the authors show that TT-cross approximation can build accurate, highly compressed representations by sampling only a fraction of the potential field. The results demonstrate compression ratios as low as 0.4% with high precision, providing a scalable "black-box" solution for complex interaction modeling.

Introduction

In mesoscale modeling, researchers bridge the gap between atomic-scale molecular dynamics and macro-scale continuum mechanics. A common approach is to treat nanostructures, such as carbon nanotubes (CNTs), as simplified geometric shapes. However, calculating the physical interactions (e.g., van der Waals forces) between these shapes is computationally expensive.

For two rigid bodies, at least six variables are required to describe their relative positions and orientations. Storing these interactions in a lookup table leads to the "curse of dimensionality," where memory requirements grow exponentially with each additional degree of freedom. While oversimplified analytical formulas are often used to circumvent this, they frequently fail to capture the nuanced physics of complex geometries. This study proposes a method to store and compute these interactions accurately and efficiently using Tensor Train (TT) decomposition.

Methodology: Tensor Train Decomposition

The authors propose using Tensor Train (TT) decomposition combined with the TT-cross approximation technique. Instead of storing a massive $d$-dimensional tensor where every entry is explicitly defined, the TT format represents the tensor as a product of small, low-rank matrices known as "cores."

The innovation lies in the "TT-cross" technique, which builds this compressed representation by sampling only a small fraction of the total possible points in the potential field. This eliminates the need to calculate the full, massive table initially. To test this, the authors modeled the van der Waals interaction between two cylindrical segments of (10,10) carbon nanotubes (radius 6.78 Å). They explored two distinct geometric parameterizations to describe the cylinders' relative positions.

Geometric parameterizations for interacting cylindrical segments: (a) one distance and three angles; (b) three distances and one angle.

Key Insights and Results

A central finding of this work is that physically meaningful functions, such as interaction potentials, are naturally "low-rank." This inherent redundancy allows the TT format to achieve remarkable compression ratios. For a grid of $40^4$ points, the TT representation compressed the data to between 0.4% and 2.4% of its original size, depending on the target precision.

The choice of coordinate system proved critical. The authors found that the first parameterization (using more angles, as seen in Figure 2a) allowed for significantly higher compression than the second. Despite this compression, the method maintained high accuracy, fully reproducing the complex "relief" of the potential energy surface at precisions of $10^{-2}$ and $10^{-3}$.

Potential energy surfaces $U$ reconstructed using the TT format as a function of angles $\alpha_1$ (a) and $\alpha_{12}$ (b) for various distances $R$.

Conclusion and Future Work

The TT-cross approach offers a "black-box" solution that is universally scalable. Unlike analytical approximations that must be hand-derived for every new shape, this method can be applied to nanotubes of varying diameters, curved segments, or even deformable bodies without altering the underlying algorithm.

While this study focused on rigid cylinders, the authors suggest that future applications could include variable geometry (length and radius) and deformability (accounting for curvature and flattening). By providing their code as open-source, the authors have enabled the broader research community to apply these data science techniques to complex physical modeling challenges.

Summary updated

Abstract

We apply the modern technique of approximation of multivariate functions – tensor train cross approximation – to the problem of the description of physical interactions between complex-shaped bodies in a context of computational nanomechanics. In this note we showcase one particular example – van der Waals interactions between two cylindrical bodies – relevant to modeling of carbon nanotube systems. The potential is viewed as a tensor (multidimensional table) which is represented in compact form with the help of tensor train decomposition. The described approach offers a universal solution for the description of van der Waals interactions between complex-shaped nanostructures and can be used within the framework of such systems of mesoscale modeling as recently emerged mesoscopic distinct element method (MDEM).
BibTeX
@article{drozdov2017time,
  author = {Grigoriy M. Drozdov and Igor A. Ostanin and Ivan V. Oseledets},
  title = {{Time- and memory-efficient representation of complex mesoscale potentials}},
  journal = {Journal of Computational Physics},
  year = {2017},
  doi = {10.1016/j.jcp.2017.04.056},
}
Figure