Implicit neural representations for chemical reaction paths

Kalyan Ramakrishnan, Lars L. Schaaf, Chen Lin, Guangrun Wang, Philip H. S. Torr

TL;DR

This paper introduces Implicit Neural Representations (INRs) as a continuous, flexible alternative to traditional discrete methods for finding the minimum energy paths (MEPs) of chemical reactions. By training a neural network to represent the entire path, their method can more robustly identify transition states, escape local minima, and even generalize to predict paths for unseen systems.

What Problem Does This Paper Address?

Predicting how chemical reactions happen and how fast they occur is crucial for fields like catalysis and drug design. A key concept is the "minimum energy path" (MEP), which describes the most likely route a reaction takes on a complex energy landscape. The highest point on this path, the "transition state" (TS), dictates the reaction's speed. Traditional methods like the Nudged Elastic Band (NEB) rely on discrete "images" or points along the path. However, NEB can struggle with complex reactions, get stuck in suboptimal paths if the initial guess is poor, or fail to capture multi-step mechanisms. This paper aims to overcome these limitations by using a continuous representation of the reaction path.

Methods

The core innovation is using Implicit Neural Representations (INRs) to model the reaction path as a smooth, continuous function. Instead of optimizing discrete points, they train a neural network whose parameters define the entire path. This allows for:

  • Continuous Parameterization: The path $x(t)$ is represented by a neural network $g_\theta(t)$ that takes a reaction coordinate $t \in [0,1]$ as input.
  • Automatic Differentiation: Leveraging the power of neural networks, they can precisely compute path tangents and gradients, which is crucial for optimization.
  • Modified Loss Function: They adapt the NEB's nudging concept by using a loss function that discards tangential energy gradients, effectively pushing the path along the valley of the potential energy surface.
  • Climbing Image Adaptation: Similar to the climbing-image NEB, they incorporate a term to push the highest-energy point on the path upwards, enabling simultaneous estimation of the MEP and TS.
  • Flexible Sampling: They introduce a "growing sampling" strategy that starts by sampling points near the reaction endpoints and gradually expands towards the center, helping to escape local minima.
  • Path Generalization: They explore training a single INR model on a dataset of reaction paths to predict MEPs for new, unseen systems, a significant step towards a universal reaction path representation.

Datasets

The paper uses several datasets for training and evaluation:

  • Two-dimensional potentials: Standard potentials like London-Eyring-Polanyi-Sato (LEPS), Müller-Brown, and custom "sine" potentials are used for initial validation and to demonstrate robustness against poor initial guesses and multi-path scenarios.
  • Atomistic systems:

    • In₂O₃ surface reaction
    • Alanine dipeptide rotation
    • Methylidyne diffusion on Ag/Cu surface
    • N₂ + H₂ reaction on a Ru surface These are challenging, high-dimensional systems where traditional NEB methods often struggle.
  • Generalization dataset: A set of 2D potentials derived from the Müller-Brown form, with varying parameters. This dataset is split into training and testing sets to evaluate the INR's ability to generalize to unseen reaction landscapes.

Key Insights and Ideas

  • Continuous representation is powerful: Moving from discrete points to a continuous neural network representation fundamentally changes how reaction paths can be optimized, offering greater flexibility and robustness.
  • Gradient manipulation is key: By carefully designing the loss function to discard tangential gradients, the INR can effectively mimic the "nudging" behavior of NEB in a continuous setting.
  • Sampling strategy matters: The "growing sampling" technique is a simple yet effective way to guide the optimization away from local minima, demonstrating that clever sampling can unlock better solutions.
  • Generalization is within reach: The ability of a single INR to learn from multiple reaction paths and predict new ones is a significant breakthrough, hinting at a future of "universal" reaction path predictors.
  • INR is more resilient to bad starting points: Unlike NEB, which can produce unphysical paths with poor initial guesses, the INR's smooth nature makes it more forgiving.

Key Results

  • Improved accuracy and robustness: The INR successfully finds MEPs and TSs for challenging atomistic systems where NEB fails or produces unphysical results (e.g., alanine dipeptide, N₂ + H₂).
  • Escape from local minima: The "growing sampling" strategy allowed the INR to find lower-barrier paths in systems with competing routes (e.g., methylidyne diffusion, sine-n potentials).
  • Efficient TS estimation: The INR can estimate the TS simultaneously with the MEP in a single optimization, often with fewer energy evaluations than NEB followed by refinement. For example, in the In₂O₃ system, the INR achieved a TS energy error comparable to refined NEB but with fewer iterations.
  • Path generalization: Trained on a limited set of 2D systems, the INR significantly outperformed nearest-neighbor and linear interpolation baselines in predicting MEPs for unseen systems, demonstrating its generalization capabilities.

Limitations and Future Work

  • Computational cost: While the INR can be more efficient for finding the TS, the initial training of the neural network can still be computationally intensive, especially for complex systems.
  • Sampling strategy optimization: The paper suggests that further improvements could be made by exploring more sophisticated sampling strategies and inductive biases in the network architecture.
  • Generalization to atomistic systems: The path generalization experiments were primarily conducted in low-dimensional settings. Extending this to high-dimensional, atomistic systems is a significant but promising future direction.
  • Internal coordinates: The authors suggest exploring the use of internal coordinates instead of Cartesian coordinates for the INR, which could improve efficiency and incorporate physical symmetries.
  • Optimizer choice: Investigating different optimizers, potentially those designed for continuous systems or ODE solvers, could further enhance performance.

Summary updated

Abstract

We show that neural networks can be optimized to represent minimum energy paths as continuous functions, offering a flexible alternative to discrete path-search methods such as Nudged Elastic Band (NEB). Our approach parameterizes reaction paths with a network trained on a loss function that discards tangential energy gradients and enables instant estimation of the transition state. We first validate the method on two-dimensional potentials and then demonstrate its advantages over NEB on challenging atomistic systems where (i) poor initial guesses yield unphysical paths, (ii) multiple competing paths exist, or (iii) the reaction follows a complex multi-step mechanism. Results highlight the versatility of the method: for instance, a simple adjustment to the sampling strategy during optimization can help escape local-minimum solutions. Finally, in a low-dimensional setting, we demonstrate that a single neural network can learn from existing paths and generalize to unseen systems, showing promise for a universal reaction path representation.
BibTeX
@article{ramakrishnan2025implicit,
  author = {Kalyan Ramakrishnan and Lars L. Schaaf and Chen Lin and Guangrun Wang and Philip H. S. Torr},
  title = {{Implicit neural representations for chemical reaction paths}},
  journal = {The Journal of Chemical Physics},
  year = {2025},
  doi = {10.1063/5.0267023},
}
Figure