Work

Potential Energy Surfaces in Multilayered Form: Deep Learning Based Regression of Analytical Hierarchical Tree Structures Using Unstructured Clouds of Data-Points

Year: 2026
Type: article
Abstract: In this work, we introduce analytical hierarchical tree structures, aka multilayered parametric functions, in the context of regression of multidimensional surfaces, such as molecular or interparticle potential energy surfaces. From a technical perspective, our binary trees stem from the ideas underlying the so-called hierarchical Tucker decomposition by replacing columnwise the factors by piecewise polynomials interpolated from randomized multivariate distributions. We refer to our approach as hierarchical Tucker finite basis representation (HT-FBR). Furthermore, our trees can be transformed into a computational graph a recursive representation in terms of operators. Our optimized HT-FBR function is smooth, continuous and differentiable. Using a single-entry rule and our direct index-flow theorem, it can be efficiently evaluated at any system configuration within the definition domain. As a consequence, we show that it is possible to optimize the HT-FBR parameters using clouds of scattered data-points in combination with a deep-learning based approach. Specifically, we show that such optimization process is very efficient when employing a backpropagation-like scheme. To assess the power of our approach, we illustrate these features in the context of multidimensional fitting of potential energy surfaces for two systems: (i) a common benchmark, a (6D) PES for the description of the cis–trans isomerization of HONO and (ii) the 12D coupled diabatic potential energy surfaces of ethene. As reference set for this last case-study, we have employed the database generated through a series of direct dynamics variational multiconfigurational Gaussian trajectories (DD-vMCG) trajectories. As a final remark, our HT-FBR representation is fully compatible with both trajectory- and wavepacket-based methods such as such (ML-)MCTDH. As a consequence, it enables a direct comparison of different nonadiabatical dynamical methods.
Language: English
Open Access status: green
OpenAlex ID: W7234134692
FWCI: 0.839
Cites: 49
Cited by: 1
Related to: 10