Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

StringNET: Neural Network based Variational Method for Transition Pathways

Jiayue Han, Shuting Gu, Xiang Zhou

TL;DR

This work addresses the computational challenge of rare transition events by reframing path finding between metastable states as a variational problem solved with neural networks. StringNET trains a single network to represent the entire transition path as a function of arc length, using a suite of loss terms to realize maximum flux, minimum energy, and minimum action paths within a unified framework; a temperature-dependent max-flux loss provides a tractable bridge to the zero-temperature MEP, and a pre-training strategy accelerates convergence for MEP and MAP. The authors demonstrate robustness and efficiency across a range of problems, from low-dimensional potentials to high-dimensional Hilbert-space functionals like the Ginzburg-Landau energy, and show that a single-network approach outperforms separate-component networks. Overall, StringNET offers a scalable, variationally grounded method for transition-path computation that is well-suited to high-dimensional systems and can be extended to energy functionals on Hilbert spaces beyond simple finite-dimensional landscapes.

Abstract

Rare transition events in meta-stable systems under noisy fluctuations are crucial for many non-equilibrium physical and chemical processes. In these processes, the primary contributions to reactive flux are predominantly near the transition pathways that connect two meta-stable states. Efficient computation of these paths is essential in computational chemistry. In this work, we examine the temperature-dependent maximum flux path, the minimum energy path, and the minimum action path at zero temperature. We propose the StringNET method for training these paths using variational formulations and deep learning techniques. Unlike traditional chain-of-state methods, StringNET directly parametrizes the paths through neural network functions, utilizing the arc-length parameter as the main input. The tasks of gradient descent and re-parametrization in the string method are unified into a single framework using loss functions to train deep neural networks. More importantly, the loss function for the maximum flux path is interpreted as a softmax approximation to the numerically challenging minimax problem of the minimum energy path. To compute the minimum energy path efficiently and robustly, we developed a pre-training strategy that includes the maximum flux path loss in the early training stage, significantly accelerating the computation of minimum energy and action paths. We demonstrate the superior performance of this method through various analytical and chemical examples, as well as the two- and four-dimensional Ginzburg-Landau functional energy.

StringNET: Neural Network based Variational Method for Transition Pathways

TL;DR

This work addresses the computational challenge of rare transition events by reframing path finding between metastable states as a variational problem solved with neural networks. StringNET trains a single network to represent the entire transition path as a function of arc length, using a suite of loss terms to realize maximum flux, minimum energy, and minimum action paths within a unified framework; a temperature-dependent max-flux loss provides a tractable bridge to the zero-temperature MEP, and a pre-training strategy accelerates convergence for MEP and MAP. The authors demonstrate robustness and efficiency across a range of problems, from low-dimensional potentials to high-dimensional Hilbert-space functionals like the Ginzburg-Landau energy, and show that a single-network approach outperforms separate-component networks. Overall, StringNET offers a scalable, variationally grounded method for transition-path computation that is well-suited to high-dimensional systems and can be extended to energy functionals on Hilbert spaces beyond simple finite-dimensional landscapes.

Abstract

Rare transition events in meta-stable systems under noisy fluctuations are crucial for many non-equilibrium physical and chemical processes. In these processes, the primary contributions to reactive flux are predominantly near the transition pathways that connect two meta-stable states. Efficient computation of these paths is essential in computational chemistry. In this work, we examine the temperature-dependent maximum flux path, the minimum energy path, and the minimum action path at zero temperature. We propose the StringNET method for training these paths using variational formulations and deep learning techniques. Unlike traditional chain-of-state methods, StringNET directly parametrizes the paths through neural network functions, utilizing the arc-length parameter as the main input. The tasks of gradient descent and re-parametrization in the string method are unified into a single framework using loss functions to train deep neural networks. More importantly, the loss function for the maximum flux path is interpreted as a softmax approximation to the numerically challenging minimax problem of the minimum energy path. To compute the minimum energy path efficiently and robustly, we developed a pre-training strategy that includes the maximum flux path loss in the early training stage, significantly accelerating the computation of minimum energy and action paths. We demonstrate the superior performance of this method through various analytical and chemical examples, as well as the two- and four-dimensional Ginzburg-Landau functional energy.
Paper Structure (24 sections, 1 theorem, 34 equations, 19 figures)

This paper contains 24 sections, 1 theorem, 34 equations, 19 figures.

Table of Contents

  1. Introduction
  2. \newlabelProblem_formulation0 Models and Path formulations
  3. Models and Notations
  4. Paths and Variational Forms
  5. MFP approximates the MEP at low temperature
  6. \newlabelmain_method0 StringNET: Numerical methods
  7. Neural network architecture
  8. Loss function for arc length parametrization
  9. \newlabelMFP_MEP0 Loss functions for max-flux path
  10. Loss function based on least squared tangent condition for minimum energy path
  11. Loss function for minimum action path in gradient systems
  12. The sum of loss functions with weights tuning, and pre-training for MEP
  13. Generalization to energy functional on Hilbert space
  14. \newlabelsec:NT0 Numerical Tests
  15. Double-well potential
  16. ...and 9 more sections

Key Result

Theorem 1

\newlabelMFP-conv-thm0 If the function $U$ is Lipschitz continuous, then the functionals $I_\beta$ Gamma converges to the point-wise limit $I_\infty$, with the topology of AC norm as $\beta\to+\infty$.

Figures (19)

  • Figure 1: One single neural network (a) used in the StringNET method. In case of the path in the function space $\varphi(s,x)$, the input is modified as $(s,x)$ where $x\in \mathbb {R}^n$ is the spatial variable and the output becomes $\mathbb {R}^1$-valued. Separate networks in (b) for each component are neither efficient in $\mathbb {R}^d$ nor applicable for PDE examples.
  • Figure 1: The initial path is the straight line between two minimizers $(\pm 1, 0)$ (dashed black). The MEP is shown in blue. The max-flux paths at different $\beta=0.35, 0.65, 1.5, 40$ are computed by the neural network function (DL) and the traditional chain-of-states method (TM). The latter method (TM) with the MATLAB subroutine fmincon fails to converge at $\beta=40$.
  • Figure 1: The numerical MEP of 2-dim Ginzburg-Landau functional ($\kappa=0.08$).
  • Figure 2: Comparison of numerical results for the MEP in the double well example. The straight line in dash is the initial. The neural network method of minimizing the loss $l_\parallel$ of mean square of tangent condition ("$\alpha_2\ell_{arc}+\alpha_3\ell_\parallel$") fails in this simple example. The other three results have the same numerical path. The minimum action method for $l_g$ ( "$\alpha_2\ell_{arc}+\alpha_4\ell_g$") works well to find the MEP. After adding the $\ell_\beta$ loss in the early stage, the method based on $l_\parallel$ can find the correct path too. Here $\alpha_2=1, \alpha_3=\alpha_4=10$ are used and $\alpha_1$, the weight for $\ell_\beta$, is set to drop from $10$ to $0$ after $2\times 10^4$ optimization steps.
  • Figure 2: The numerical MEP of 2-dim Ginzburg-Landau functional ($\kappa=0.05$) obtained by using the $\ell_\beta$ pre-training.
  • ...and 14 more figures

Theorems & Definitions (2)

  • Theorem 1
  • Proof 1