Boosting Computational Catalysis and Chemical Reactivity with Artificial Intelligence

Konstantinos D. Vogiatzis, Clémence Corminboeuf, Ainara Nova, Kjell Jorner, Johannes Kästner, Markus Meuwly, Philippe Schwaller, Victor Böttcher, Maria Drosou, Edvin Fako, Hannes Hoppe, Žarko Ivković, Nestor M. Iwanojko, Dimitrios A. Pantazis, Stefan P. Schmid, Kalman Szenes, Auguste Tetenoire, Markus ReiherJournal of the American Chemical Society 2026 · DOI

TL;DR

Artificial intelligence (AI) and machine learning (ML) are revolutionizing computational chemistry, particularly in catalysis. This perspective highlights how AI/ML, from machine learning potentials to generative models and large language models, can accelerate catalyst discovery and deepen our understanding of chemical reactivity, while also emphasizing the crucial role of human expertise and robust data.

What Problem Does This Paper Address?

The field of computational catalysis, especially involving transition metal complexes, has traditionally been computationally intensive and slow. Discovering new catalysts and understanding complex reaction mechanisms requires significant human effort and expertise. While AI/ML has made strides in organic chemistry, its application to transition metal chemistry and catalysis is still developing. This paper addresses the challenges and opportunities in leveraging AI/ML to overcome these limitations, aiming to accelerate discovery and enhance understanding in this critical area of chemistry.

Methods

This perspective focuses on several emerging AI/ML methodologies:

  • Machine Learning Interatomic Potentials (MLIPs): These models aim to replace traditional quantum chemical methods (like DFT) for molecular simulations, offering significant speedups while maintaining or even surpassing accuracy. They are trained on data from more rigorous methods and are becoming increasingly generalizable.
  • Generative AI: Techniques like diffusion and flow-matching models are being used to directly generate 3D molecular geometries, enabling inverse design of catalysts and exploration of chemical space.
  • Foundation Models (including Large Language Models - LLMs): These models, trained on vast datasets, aim for broad applicability. In chemistry, they can serve as universal energy and force predictors (MLIPs) or process and reason about chemical information (LLMs), aiding in tasks from coding to scientific ideation.
  • Transfer Learning and Active Learning: These techniques are crucial for efficiently training models, especially in data-scarce domains, and for intelligently guiding the exploration of complex chemical spaces.

Datasets

The paper emphasizes the critical need for high-quality, diverse datasets. While specific datasets used for training MLIPs or generative models are not detailed, the authors highlight:

  • Scarcity of Transition Metal Complex Data: A significant challenge is the lack of comprehensive, computationally refined databases for transition metal complexes.
  • Importance of Negative Results: The absence of published "failed" reactions creates a positive bias in training data, hindering models' ability to learn what doesn't work. The authors advocate for incentivizing the publication of negative outcomes.
  • Electronic Lab Notebooks and Open Databases: These are proposed as solutions for systematic data storage and sharing, facilitating the creation of more robust and interoperable datasets.
  • OMOL2025 Dataset: Mentioned as an example of a general-purpose dataset for training foundation models.

Key Insights and Ideas

  • AI as an Accelerator, Not a Replacement: The core message is that AI/ML should augment, not replace, human intuition and expertise in chemistry. It's a powerful tool for accelerating discovery and insight.
  • Bridging the "Model-Mechanism Gap": AI/ML can help connect theoretical predictions with experimental observations, particularly by correlating spectroscopic data with molecular/electronic structure.
  • The Power of Representations: The way chemical information is encoded (molecular representations) is paramount. Multimodal representations that integrate various data types (text, graphs, images, spectra) are seen as a promising future direction.
  • Domain Knowledge is Still King: For complex or under-characterized systems (like transition metal complexes), hand-crafted features and expert knowledge remain crucial for guiding AI/ML models.
  • Caution Against "Panaceas": Just as the Universal Force Field (UFF) had limitations, the pursuit of universally applicable AI models needs to be balanced with the demand for rigor and specificity in chemistry.
  • The Rise of Self-Driving Laboratories: The integration of AI/ML with automation and robotics points towards future autonomous discovery platforms.

Key Results

The paper is a perspective, not a research article reporting specific quantitative results. However, it highlights the potential of AI/ML:

  • Accelerated Catalyst Discovery: MLIPs and generative models can drastically speed up the search for new catalysts.
  • Improved Mechanistic Understanding: AI can help analyze complex reaction pathways and correlate experimental observables with theoretical structures.
  • Enhanced Data Utilization: Techniques like transfer learning allow models to be effective even with limited data.
  • Potential for Automation: AI/ML is paving the way for automated hypothesis generation, experimental design, and iterative refinement in research.

Limitations and Future Work

  • Robust Molecular Representations for Transition Metals: Developing universal and accurate representations for complex transition metal systems remains a challenge.
  • Bridging Mechanistic Understanding with AI: Integrating deep chemical intuition and mechanistic knowledge into AI workflows is an ongoing effort.
  • Reliable and Diverse Datasets: The scarcity of data, especially negative results, is a major bottleneck.
  • Interpretability vs. Performance: Balancing the need for models that are accurate and generalizable with those that are understandable and explainable is crucial.
  • Extrapolation to Underexplored Regions: MLIPs can sometimes produce unreliable results in regions of chemical space not covered by training data.
  • Future Work: The authors suggest continued development of MLIPs, generative models, LLMs, and multimodal approaches. They emphasize the importance of integrating human expertise, creating better datasets, and developing more interpretable AI tools. The ultimate goal is to create AI systems that act as powerful collaborators for chemists, accelerating discovery and deepening understanding.

Summary updated

Abstract

Artificial intelligence (AI) and machine learning (ML) are rapidly reshaping the landscape of computational chemistry, offering new opportunities for accelerating catalyst discovery and deepening our understanding of chemical reactivity. This perspective highlights emerging methodologies ranging from machine learning potentials and reinforcement learning to generative AI and large language models that are poised to transform computational catalysis. We discuss challenges in developing robust molecular representations for transition-metal complexes, bridging mechanistic understanding with AI-driven predictions, and constructing reliable data sets that capture both successful and failed reactivity outcomes. By drawing on the authors' practical experience across computational, experimental, and AI-driven domains, we emphasize the importance of integrating chemical intuition and methodological expertise with data-driven approaches while remaining open to serendipitous discoveries enabled by automation and self-driving laboratories. Ultimately, the future of computational catalysis lies in balancing human intuition with algorithmic power, leveraging AI not as a replacement but as an accelerator of chemical insight, mechanistic understanding, and catalyst design.
BibTeX
@article{vogiatzis2026boosting,
  author = {Konstantinos D. Vogiatzis and Clémence Corminboeuf and Ainara Nova and Kjell Jorner and Johannes Kästner and Markus Meuwly and Philippe Schwaller and Victor Böttcher and Maria Drosou and Edvin Fako and Hannes Hoppe and Žarko Ivković and Nestor M. Iwanojko and Dimitrios A. Pantazis and Stefan P. Schmid and Kalman Szenes and Auguste Tetenoire and Markus Reiher},
  title = {{Boosting Computational Catalysis and Chemical Reactivity with Artificial Intelligence}},
  journal = {Journal of the American Chemical Society},
  year = {2026},
  doi = {10.1021/jacs.5c17786},
}
Figure