Combining machine learning methods and reactivity theories to better understand and analyze complex chemical processes

Molecular quantum mechanical modeling, accelerated by machine learning, has opened the door to high-throughput screening campaigns of chemical reactions, enabling the computational discovery of new catalysts, new bioorthogonal click reactions and new self-healing polymers. In this research line, we develop tools, approaches and strategies to facilitate the set-up of these types of reaction discovery workflows, and we demonstrate how they can be put to productive use.

Selected publications:

1. N. Casetti, J. E.. Alfonso‐Ramos, C. W. Coley, T. Stuyver, Combining molecular quantum mechanical modeling and machine learning for accelerated reaction screening and discovery, Chem. Eur. J. 2023, e202301957.
2. M. Ferrer, B. Deng, J. Alfonso-Ramos, T. Stuyver, Screening Diels-Alder reaction space to identify candidate reactions for self-healing polymer applications. Digit. Discov. 2025, 4, 3400-3410.
3. T. Stuyver, C. Coley, Machine Learning‐Guided Computational Screening of New Candidate Reactions with High Bioorthogonal Click Potential, Chem. Eur. J. 2023, e202300387.

Reaction optimization often requires exploring many combinations of catalysts, ligands, solvents, and experimental conditions. Bayesian optimization uses predictive models and their uncertainty estimates to select experiments sequentially, aiming to identify suitable conditions with as few measurements as possible. In this research line, we study how molecular representations interact with Gaussian-process models, and how observations from previous campaigns can inform the optimization of related systems. Alongside these methodological developments, we also collaborate with experimentalists on prospective optimization campaigns.

Selected publications:

1. G. Chen, M. Fleck, T. Stuyver, Robust transfer learning for Bayesian optimization of chemical reactions, Digit. Discov. 2026, DOI: 10.1039/d6dd00527f.
2. G. Chen, M. Fleck, T. Stuyver, Leveraging hidden-space representations effectively in Bayesian optimization for experiment design through dimension-aware hyperpriors, J. Chem. Theory Comput. 2026, 22, 5594-5608.
3. B. Zhang, H. Zeng, H. Yin, D. Zheng, Z. Wan, C. Jia, T. Stuyver, J. Luo, T. Pauporté, Combining component screening, machine learning and molecular engineering for the design of high-performance inverted perovskite solar cells, Energy Environ. Sci. 2024, 17, 5532-5541.

Traditional machine learning algorithms tend to require lots of training data. However, by introducing domain knowledge, one can dramatically improve the data efficiency -- as well as generalizability -- of these models. In this research line, we aim to augment models with descriptors stemming from quantum chemical calculations to facilitate their application to domains for which only limited data is available.

Selected publications:

1. G. Chen, T. Stuyver, Harnessing Surrogate Models for Data-efficient Predictive Chemistry: Descriptors vs. Learned Hidden Representations, Digit. Discov. 2025, 4, 3227-3237.
2. T. Stuyver, C. Coley, Quantum chemistry-augmented neural networks for reactivity prediction: Performance, generalizability, and explainability, J. Chem. Phys. 2022, 156, 084104.
3. J. E. Alfonso-Ramos, R. Neeser, T. Stuyver, Repurposing quantum chemical descriptor datasets for on-the-fly generation of informative reaction representations: application to hydrogen atom transfer reactions, Digit. Discov. 2024, 3, 919-931. 

Many processes studied in transition-metal catalysis, bio- and polymer chemistry, enzyme catalysis and environmental chemistry involve a multitude of elementary reaction steps and competing reaction pathways. In order to analyze, control and/or steer the outcome of generic, complex reaction networks, all the relevant chemical compounds and associated elementary reactions need to be identified. In this research line, we aim to combine efficient algorithms and machine learning approaches to efficiently explore and analyze such networks.

Selected publications:

1. Z. Tu, T. Stuyver, C. Coley, Predictive chemistry: machine learning for reaction deployment, reaction development, and reaction discovery, Chem. Sci. 2023, 14, 226-244.
2. T. Stuyver, TS‐tools: Rapid and automated localization of transition states based on a textual reaction SMILES input, J. Comput. Chem. 2024, 45, 2308-2317
3. J. E. Alfonso-Ramos, T. Stuyver, Reaction enumeration based on NBO-informed molecular graphs, J. Comput. Chem. 2026, 47, e70462. 

Effective QM augmentation of machine learning models inherently requires a thorough understanding of the fundamental principles of chemical reactivity. In this research line, we aim to expand and unify existing reactivity theories (conceptual Density Functional Theory, Valence Bond Theory, Molecular Orbital Theory etc.) through consideration of various reactivity problems.

Selected publications:

1. T. Stuyver, F. De Proft, P. Geerlings, S. Shaik, How do local reactivity descriptors shape the potential energy surface associated with chemical reactions? The valence bond delocalization perspective, J. Am. Chem. Soc. 2020, 142, 10102-10113.
2. T. Stuyver, S. Shaik, Unifying conceptual density functional and valence bond theory: The hardness–softness conundrum associated with protonation reactions and uncovering complementary reactivity modes, J. Am. Chem. Soc. 2020, 142, 20002-20013.
3. T. Stuyver, B. Chen, T. Zeng, P. Geerlings, F. De Proft, R. Hoffmann, Do diradicals behave like radicals?, Chem. Rev. 2019, 119, 11291-11351.