What we work on
We develop machine learning methods to understand and predict chemical reactions. Our work spans the full pipeline from data generation and curation to model development and benchmarking, bridging data-driven AI and mechanistic chemistry.
Our vision
Machine learning for chemical reactions has the potential to enable high-throughput screening at a scale previously out of reach, accelerating discovery in drug synthesis, materials design, and sustainable chemistry.
The current challenges in the field span representation, model architecture, and data:
We believe ML models should be built around chemical principles from the ground up: starting with meaningfully curated datasets, moving to chemically informed representations, and reaction-specific architectures. Together, these lay the foundation for direct, reliable prediction of reaction energetics at the scale real applications demand.
Areas of research
Working toward this goal, we tackle every step along the way: from the models themselves, through the automated pipelines that put them to work, to the data everything depends on.
01 Prediction
Predicting properties of chemical reactions, such as barrier heights or reaction enthalpies, is a key task in computational chemistry. Machine learning offers a powerful framework for approximating reaction properties, without the cost of exhaustive quantum mechanical calculations. As a foundation for community-driven method development and reproducible benchmarking, we developed ChemTorch, an open-source deep learning framework that streamlines model development and benchmarking across different modalities for chemical reaction property prediction.
A particular focus of our work lies in the prediction of reaction barrier heights, which govern reaction rates and selectivity and are notoriously expensive to obtain via traditional QM methods. We have shown that graph neural networks operating on the condensed graph of reaction (CGR) provide high predictive accuracy, and that further incorporating transition state geometries significantly improves barrier height predictions over purely connectivity-based approaches.
Beyond graph-based methods, we also investigate language model approaches for reaction property prediction. We introduced sr-SMILES, a compact sequence-based representation derived from the CGR that encodes explicit mechanistic and stereochemical information in a textual format amenable to transformer models.
02 Generative models
Generative models have emerged as a promising approach for transition state (TS) geometry generation. Our method GoFlow frames TS generation as an optimal transport flow problem, solved via E(3)-equivariant flow matching with geometric tensor networks. Compared to prior diffusion-based approaches, GoFlow achieves a significant speedup in inference while improving geometric accuracy, enabling deep learning-based TS generation for high-throughput computational chemistry workflows.
While models like GoFlow generate TS geometries with high accuracy, they lack chirality awareness. We address this with Feynman-Kac-Flow, which derives Feynman-Kac steering for conditional flow matching. By tilting the generated distribution with an energy potential, our approach steers samples toward chirality-consistent transition states, marking an important step toward chirality-aware TS generation.
Beyond chemical reactions, the GoFlow framework generalizes to other geometry prediction tasks, such as minimum-energy triplet geometries. GoFlow-T1 adapts the same backbone to predict these excited-state geometries directly across chemically diverse organic molecules, enabling rapid computation of triplet energies and singlet-triplet gaps at near-quantum-mechanical accuracy.
03 Automated workflows
Generating accurate TS geometries is only one part of the challenge; integrating them into robust, scalable workflows for reaction mechanism exploration is another. Standalone deep learning predictors lack robustness under distribution shift, while heuristic-based automation struggles to scale, limiting their utility for high-throughput discovery.
With moTSart we address this by embedding a generative model directly within a fully automated TS search pipeline, coupled with physics-based validation. Rather than replacing quantum mechanics, the model accelerates it by proposing high-quality geometry candidates that reduce the number of costly optimization cycles required. The framework operates in a continuous learning regime, improving as more reactions are processed: in a high-throughput screening study of bio-orthogonal click reactions, the required TS optimization cycles dropped by nearly 50% after processing just 540 reactions.
04 Data
While ML models and representations for chemical reactions have matured rapidly, data is a real bottleneck. Existing datasets tend to be small, restricted to gas-phase reactions, and often mechanistically incomplete. We believe that progress depends on generating the right data: mechanistically diverse, internally consistent, and grounded in chemically meaningful elementary steps.
We develop principled strategies for data curation and generation — prioritizing chemical diversity, consistent levels of theory, and validated transition states over sheer dataset size.
We work in the wet lab to generate reaction data ourselves. A current focus is asymmetric organocatalysis, where enantioselectivity serves as a direct and measurable experimental observable, and reaction spaces grow combinatorially, making data-driven, intelligent sampling of the reaction space essential.
Funding
A big part of our research is supported by a €1.5M ERC Starting Grant, awarded to Esther Heid in 2024 to develop data-driven deep-learning frameworks for organic and enzymatic reactions.
Explore the ERC project →
We are further supported by the following grants and programs:
Open source
We hold our work to a high scientific standard and share it openly. Our models, code, and datasets are released as open source, so the community can reproduce our results and build on them.