CatTSunami: Accelerating Transition State Energy Calculations with Pre-trained Graph Neural Networks

Brook Wander, Muhammed Shuaibi, John R. Kitchin, Zachary W. Ulissi, C. Lawrence Zitnick

TL;DR

This paper introduces CatTSunami, a framework that leverages pre-trained graph neural networks (GNNs) to dramatically accelerate the calculation of transition state energies for chemical reactions. By achieving 91% accuracy with a 28x speedup compared to traditional methods, this approach promises to make large-scale catalyst discovery and reaction mechanism analysis computationally feasible.

What Problem Does This Paper Address?

Discovering new catalysts is crucial for developing sustainable chemical processes, but it's a slow, trial-and-error endeavor. A major bottleneck is the computational cost of accurately determining transition state energies, which are essential for understanding reaction rates and mechanisms. Traditional methods like Density Functional Theory (DFT) coupled with the Nudged Elastic Band (NEB) method are highly accurate but prohibitively expensive for exploring the vast chemical space of potential catalysts and reaction pathways. This limits the scope and depth of computational studies, often forcing researchers to rely on approximations that may not hold universally.

Methods

The core of this work is the application of pre-trained Graph Neural Networks (GNNs) – specifically, models trained on the Open Catalyst 2020 (OC20) dataset for local adsorbate relaxations – to the task of calculating transition state energies via NEB. This is a "zero-shot" application, meaning the models were not explicitly trained on reaction data. The authors explored four different strategies for integrating ML into the NEB workflow:

  • All ML: ML models are used for all force and energy evaluations.
  • ML + 3 DFT SPs: ML models relax initial/final frames, and DFT single points are used for initial, transition, and final states.
  • ML + 2 DFT RXs + 1 DFT SP: Initial/final frames are relaxed with DFT, then ML is used for the NEB calculation, with a DFT single point on the transition state. This was the primary method for comparison.
  • ML pre-relaxation + DFT RX NEB: Initial/final frames are relaxed with DFT, and the NEB calculation itself is pre-relaxed with ML before a final DFT relaxation.

The CatTSunami framework automates the generation of initial and final frames for NEB calculations, leveraging ML for local relaxations and proximity criteria to propose reasonable starting points.

Datasets

  • OC20 Dataset: Used for pre-training the GNN models. This dataset consists of ~260 million DFT single-point calculations on local adsorbate relaxations across 82 adsorbates and 12,000 materials.
  • OC20NEB Dataset: A newly curated dataset introduced in this paper, comprising 932 revised Perdew-Burke-Ernzerhof (RPBE) DFT NEB calculations. This dataset includes desorption, dissociation, and transfer reactions on diverse, randomly generated surfaces, with half of the materials being "out-of-domain" relative to the OC20 training set.

Key Insights and Ideas

  • Generalizability of Pre-trained Models: The most striking insight is that GNNs pre-trained on local adsorbate relaxations (a related but distinct task) can perform remarkably well on transition state energy calculations without any further training. This highlights the power of transferable representations learned by these models.
  • Significant Computational Speedup: The ability to achieve DFT-level accuracy with a substantial speedup (up to 1500x in one case) is transformative. This opens the door to exhaustively exploring reaction networks that were previously intractable.
  • Beyond BEP Approximations: The work demonstrates that ML-accelerated NEB can provide more accurate and reliable kinetic data than traditional Brønsted-Evans-Polanyi (BEP) relations, especially for complex reaction networks with diverse reactions.
  • Discovery of Lower Energy Configurations: In case studies, the ML approach not only accelerated calculations but also systematically found lower energy transition states and intermediates compared to previous studies, suggesting a more thorough exploration of the potential energy surface.

Key Results

  • High Success Rate: The top-performing ML models achieved an average success rate of 91% in finding transition states energetically similar (within 0.1 eV) to DFT calculations.
  • Significant Speedup: A 28x speedup was observed for the ML + 2 DFT RX + 1 DFT SP approach, with a 1500x speedup achieved for a large reaction network using the "All ML" approach (12 GPU days vs. 52 GPU years).
  • Reduced MAE: For the CO hydrogenation network, ML-accelerated NEBs achieved a Mean Absolute Error (MAE) of 0.04 eV, a significant improvement over BEP relations (MAE of 0.20 eV).
  • Structural Similarity: Analysis of reaction coordinates showed strong agreement between ML and DFT determined transition state structures, indicating that the ML models capture not just energy but also the geometry of the reaction pathway.

Limitations and Future Work

  • "Zero-Shot" Limitations: While impressive, the zero-shot performance might not be optimal for all systems. The authors note that some models showed slightly lower convergence rates for specific reaction types, and further refinement might be needed for certain complex cases.
  • Spin Polarization: The pre-trained models do not inherently account for spin polarization, which required additional refinement for certain elements (Co and Ni) in the ammonia synthesis case study.
  • Reaction Coordinate Similarity: While energy comparisons were excellent, some failure cases showed significant differences in reaction coordinates, raising questions about whether ML found incorrect or simply different valid transition states.
  • Future Work: The authors suggest that this approach can be extended to other NEB alternatives (e.g., dimer method, growing string method) and that the framework can be used for more comprehensive catalyst screening and mechanism elucidation, potentially changing how detailed kinetic studies are performed. The development of more robust automated enumeration and optimization schemes is also implied.

Summary updated

Abstract

Direct access to transition state energies at low computational cost unlocks the possibility of accelerating catalyst discovery. We show that the top performing graph neural network potential trained on the OC20 dataset, a related but different task, is able to find transition states energetically similar (within 0.1 eV) to density functional theory (DFT) 91% of the time with a 28x speedup. This speaks to the generalizability of the models, having never been explicitly trained on reactions, the machine learned potential approximates the potential energy surface well enough to be performant for this auxiliary task. We introduce the Open Catalyst 2020 Nudged Elastic Band (OC20NEB) dataset, which is made of 932 DFT nudged elastic band calculations, to benchmark machine learned model performance on transition state energies. To demonstrate the efficacy of this approach, we replicated a well-known, large reaction network with 61 intermediates and 174 dissociation reactions at DFT resolution (40 meV). In this case of dense NEB enumeration, we realize even more computational cost savings and used just 12 GPU days of compute, where DFT would have taken 52 GPU years, a 1500x speedup. Similar searches for complete reaction networks could become routine using the approach presented here. Finally, we replicated an ammonia synthesis activity volcano and systematically found lower energy configurations of the transition states and intermediates on six stepped unary surfaces. This scalable approach offers a more complete treatment of configurational space to improve and accelerate catalyst discovery.
BibTeX
@inproceedings{wander2025cattsunami,
  author = {Brook Wander and Muhammed Shuaibi and John R. Kitchin and Zachary W. Ulissi and C. Lawrence Zitnick},
  title = {{CatTSunami: Accelerating Transition State Energy Calculations with Pre-trained Graph Neural Networks}},
  booktitle = {ACS},
  year = {2025},
  doi = {10.1021/acscatal.4c04272},
}
Figure