Model Order Reduction for Quantum Molecular Dynamics
Siu Wun Cheung, Youngsoo Choi, Jean-Luc Fattebert, Daniel Osei-Kuffuor · 2025
Read Model Order Reduction for Quantum Molecular Dynamics by Siu Wun Cheung, Youngsoo Choi, Jean-Luc Fattebert, Daniel Osei-Kuffuor (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-07340.txt.
Introduction
Simulating molecules from quantum principles is accurate but slow, because each step needs a costly electronic calculation. Siu Wun Cheung and 3 co-authors offer a shortcut in a 2025 research paper. Their model order reduction method learns a compact picture of the electronic solution space. It then finds the ground state density, the lowest-energy electron arrangement, inside that smaller space.
By skipping the full iterative search over electron wavefunctions, the method cuts computing cost sharply. The team tests it on a water molecule. Results match high-fidelity simulations closely for both molecular shape and dynamic behavior. Careful parametrization and systematic sampling help the reduced model stay accurate beyond its training cases. Faster quantum dynamics could widen studies of reactions and materials that are now too expensive to simulate.
Catalog record
| CREATOR | Siu Wun Cheung, Youngsoo Choi, Jean-Luc Fattebert, Daniel Osei-Kuffuor |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 16 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-07340
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Model Order Reduction for Quantum Molecular Dynamics Siu Wun Cheung, Youngsoo Choi, Jean-Luc Fattebert, Daniel Osei-Kuffuor Abstract Molecular dynamics simulations are indispensable for exploring the behavior of atoms and molecules. Grounded in quantum mechanical principles, quantum molecular dynamics provides high predictive power but its computational cost is dominated by iterative high-fidelity electronic structure calculations. We propose a novel model order reduction approach as an alternative to high-fidelity electronic structure calculation. By learning a low-dimensional representation of the electronic solution manifold within the Kohn-Sham density functional theory framework, our model order reduction approach determines the ground state electronic density by projecting the problem onto a low-dimensional subspace, thereby avoiding the computationally expensive iterative optimization of electronic wavefunctions in the full space. We demonstrate the capability of our method on a water molecule, showing excellent agreement with high-fidelity simulations for both molecular geometry and dynamic properties, highlighting the generalizability through carefully designed parametrization and systematic sampling. \workshoptitle Machine Learning and the Physical Sciences Model Order Reduction for Quantum Molecular Dynamics Your Name(s) Your Affiliation(s) 1 Introduction Molecular dynamics (MD) frenkel2023understanding is a crucial tool in modern science for understanding the behavior of atoms and molecules over time, providing insights into material properties, chemical reactions, and biological processes. A rapidly growing area in machine learning applied to molecular dynamics is the development of machine learning interatomic potentials (MLIPs) deringer2019machine . These methods train neural networks or other machine learning models to directly map atomic configurations to energies and forces, effectively replacing the computationally intensive quantum mechanical calculation with a cheaper surrogate model. While MLIPs have been highly successful in achieving significant speed-ups, their primary limitation lies in their reliance on fitting a potential energy surface. The forces derived from MLIPs do not necessarily originate from a direct minimization of a fundamental quantum mechanical energy functional at each step. This can compromise their scientific consistency and transferability , particularly when extrapolating to novel chemical environments or extreme conditions not well-represented in their training data. In contrast, quantum molecular dynamics (QMD) marx2009ab provides a fundamental framework for simulating the evolution of atoms and molecules with high predictive power, based directly on quantum mechanical principles without relying on empirical potentials. However, its core computational bottleneck lies in the need to solve a complex, nonlinear electronic structure eigenvalue problem, typically within the Kohn-Sham (KS) Density Functional Theory (DFT) framework cances2023density , at every simulation timestep. This computational expense severely limits the scale and duration of QMD simulations, hindering its application to many important scientific problems where long timescales or large systems are crucial. In this work, we propose a model order reduction (MOR) approach to avoid repeated expensive computations in electronic structures while maintaining the rigorous physical foundation of QMD. Employing projection-based reduction, which have been successfully applied to many large-scale and complex scientific problems choi2021space ; brunini2022projection ; copeland2022reduced ; cheung2023local ; tsai2023accelerating ; chung2024scalable ; rathore2025projection ; youkilis2025projection , our method aims to determine the ground state electronic density in low-dimensional structures within the DFT framework, which ensures a higher degree of scientific consistency and predictive reliability compared to purely empirical MLIPs, while crucially avoiding expensive iterative wavefunction optimization. This new approach involves learning a compact representation of the electronic solution manifold and projecting the electronic structure problem onto this low-dimensional subspace, bypassing the need for computationally intensive wavefunction optimization at each MD step. While generalization to unseen configurations is often non-trivial in data-driven models, we demonstrate that, with carefully designed parametrization and systematic sampling strategies, the generalizability of our model can be guaranteed, leading to robust and accurate MD predictions. 2 Quantum Molecular Dynamics QMD offers a powerful and fundamental approach to simulating the coupled evolution of atomic nuclei and electrons. Central to QMD is the Born-Oppenheimer approximation, which assumes that due to their much larger mass, atomic nuclei move significantly slower than electrons. This allows the electronic ground state to be determined at each instantaneous nuclear configuration, which then provides the forces that drive the nuclear motion. The temporal evolution of the ionic positions \{\mathbf{R}_{I}\}_{I=1}^{N_{I}} is typically integrated using schemes such as the Verlet algorithm: \mathbf{R}_{I}(t+\Delta t)=2\mathbf{R}_{I}(t)-\mathbf{R}_{I}(t-\Delta t)+\frac{\Delta t^{2}}{M_{I}}\mathbf{F}_{I}(t), (1) where M_{I} is the mass of ion I and \mathbf{F}_{I}(t) is the total force acting on it. The force driving nuclear motion is derived from the electronic ground state energy, and is given by the derivative of the system’s potential energy surface with respect to the nuclear coordinates: \mathbf{F}_{I}=-\nabla_{\mathbf{R}_{I}}E[\{\phi_{i}\}_{i=1}^{N}], (2) where E[\{\phi_{i}\}_{i=1}^{N}] is the total electronic energy functional, which is a functional of the ground state electron wavefunctions \{\phi_{i}\}_{i=1}^{N} at each time t . This derivative is often computed using the Hellmann-Feynman theorem, which states that if the electronic wavefunctions are exact eigenfunctions of the Hamiltonian, the forces can be calculated as the expectation value of the gradient of the potential energy with respect to the nuclear coordinates. The accurate determination of these forces fundamentally relies on solving the electronic structure problem, typically within the framework of KS DFT. The core of KS DFT involves solving a set of effective single-particle equations: H\phi_{i}=\epsilon_{i}\phi_{i}, (3) where H is the KS Hamiltonian, \phi_{i} are the KS orbitals, and \epsilon_{i} are their corresponding energies. The Hamiltonian H itself depends on the ionic positions \{\mathbf{R}_{I}\}_{I=1}^{N_{I}} as well as the electron density \rho_{e}(\mathbf{r})=2\sum_{i=1}^{N}f_{i}|\phi_{i}(\mathbf{r})|^{2} , where N is the number of electronic states and f_{i}\in[0,1] are the occupation factors, typically given by the Fermi-Dirac distribution. This hierarchical structure underscores the significant computational complexity inherent in QMD simulations: an outermost loop for time integration, an inner loop for the nonlinear DFT self-consistent field (SCF) iterations, and within each SCF iteration, an eigenvalue problem solution to update the KS orbitals. 3 Model Order Reduction The primary computational challenge in QMD stems from the iterative, nonlinear solution of the KS equations at every MD timestep. For systems with a large number of atoms or when high accuracy is required, this step becomes exceptionally expensive, severely limiting the feasible simulation size and duration. To address this issue, we propose a MOR approach. Our method leverages a pre-computed, low-dimensional subspace to approximate the electronic structure, thereby enabling the determination of the electronic density without a full, iterative wavefunction optimization in a high-dimensional search space. The construction of this effective reduced basis is crucial. For molecular systems, we exploit the inherent equivariance of the electronic structure with respect to rigid body motions (translations and rotations) of the entire molecule. By transforming all atomic configurations into a canonical reference frame, the parametric domain for training data can be significantly simplified. For instance, by pinning the oxygen atom of a water molecule at the origin and aligning its principal axes, only a few intrinsic parameters, such as bond lengths and bond angles, characterize the molecular configurations. This reduction in the effective dimensionality of the configuration space allows for more efficient and comprehensive sampling of the relevant electronic states with a much smaller number of training points. This approach makes data collection for the reduced basis tractable and enables robust predictions for unseen configurations in molecular dynamics simulations. The MOR workflow is divided into offline and online stages: Offline Stage: As shown on the left of Figure 1 , we sample the parametric domain to generate training data.