Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

sbml4md: A computational platform for System-Bath Modeling via Molecular Dynamics powered by Machine Learning

Kwanghee Park, Seiji Ueno, Yoshitaka Tanimura

Abstract

We introduce sbml4md, a newly developed algorithm implemented as a software package to extract parameters of multimode anharmonic Brownian (MAB) models from molecular dynamics (MD) trajectories for simulating nonlinear vibrational spectra of intramolecular modes of molecular liquids. By leveraging machine learning (ML) techniques to capture vibrational anharmonicity, intermolecular couplings, and bath correlation functions for each mode, sbml4md obviates empirical fitting and enables the modeling of environments with spatial and temporal heterogeneity. This work provides a set of parameters specifically tailored for the Hierarchical Equations of Motion (HEOM) framework, enabling numerically "exact" simulations of nonlinear vibrational spectra. Building upon our previous implementation for intramolecular vibrational modes [Park, Jo, and Tanimura, J. Chem. Phys. 163, 214104 (2025)], the present code enhances optimization efficiency by explicitly accounting for intermolecular vibrational contributions. This extension enables sbml4md to broaden the applicability of HEOM-based dynamical modeling by seamlessly integrating classical MD approaches, thereby providing a flexible and scalable framework for simulating both linear and nonlinear spectra under realistic conditions with minimal empirical input. The accompanying ML code, written in Python, is provided as supporting material.

sbml4md: A computational platform for System-Bath Modeling via Molecular Dynamics powered by Machine Learning

Abstract

We introduce sbml4md, a newly developed algorithm implemented as a software package to extract parameters of multimode anharmonic Brownian (MAB) models from molecular dynamics (MD) trajectories for simulating nonlinear vibrational spectra of intramolecular modes of molecular liquids. By leveraging machine learning (ML) techniques to capture vibrational anharmonicity, intermolecular couplings, and bath correlation functions for each mode, sbml4md obviates empirical fitting and enables the modeling of environments with spatial and temporal heterogeneity. This work provides a set of parameters specifically tailored for the Hierarchical Equations of Motion (HEOM) framework, enabling numerically "exact" simulations of nonlinear vibrational spectra. Building upon our previous implementation for intramolecular vibrational modes [Park, Jo, and Tanimura, J. Chem. Phys. 163, 214104 (2025)], the present code enhances optimization efficiency by explicitly accounting for intermolecular vibrational contributions. This extension enables sbml4md to broaden the applicability of HEOM-based dynamical modeling by seamlessly integrating classical MD approaches, thereby providing a flexible and scalable framework for simulating both linear and nonlinear spectra under realistic conditions with minimal empirical input. The accompanying ML code, written in Python, is provided as supporting material.
Paper Structure (35 sections, 32 equations, 6 figures, 4 tables)

This paper contains 35 sections, 32 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. MAB Model
  3. sbml4md: Package Overview
  4. physics/---Baths, Potentials, Couplings:
  5. model/---S-B Composition and Rollout:
  6. training/---Keras Wrapper, Samplers, and Datasets:
  7. integrators/---Time Propagation:
  8. config/---YAML Schemas and Examples:
  9. execution graph: Training and Inference:
  10. Training Pipeline and Model I/O Specification
  11. Minimal Input Requirements
  12. Rollout Model Interface
  13. Dataset Construction and Sampling
  14. Configuration-Driven Entry Point
  15. The core steps
  16. ...and 20 more sections

Figures (6)

  • Figure 1: Software module architecture.
  • Figure 2: Physics module diagram.
  • Figure 3: Infrared absorption spectra evaluated using (a) the flexible SPC/E potential and (b) the Ferguson potential for water are shown. The spectra were calculated with CHFPE using MAB parameters trained on MD trajectories generated with the flexible SPC/E and Ferguson potentials. Both the stretching and bending vibrational modes are clearly observed. For comparison, each panel also includes results from MD simulations (blue lines) and experimental data (black dashed curves).IRexp2011 The experimental spectrum is reproduced with permission from Author, J. Mol. Struct. 1004, 146 (2011). Copyright (2011) Elsevier.
  • Figure 4: PBM spectra in the low-frequency region for the (a) flexible SPC/E potential and (b) Ferguson potential. Here, PBM1 (purple), PBM2 (green), and PBM3 (magenta) are plotted using the fitted PBM parameters listed in Table \ref{['tab:spce_drude_3bo_compact_products']}. For comparison, each panel also includes results from MD simulations (blue lines) and experimental data (black dashed curves).IRexp2011 The experimental spectrum is reproduced with permission from Author, J. Mol. Struct. 1004, 146 (2011). Copyright (2011) Elsevier.
  • Figure 5: 2D correlation IR spectra for the stretching and stretching$\rightarrow$ bending motions were obtained using optimized parameter values from (a) the flexible SPC/E model and (b) the Ferguson model. Because the peak intensities in the lower panels are weaker than those in the upper panels, the contour interval was tripled to enhance visibility.
  • ...and 1 more figures