| _version_ | 1866901694395187200 |
|---|---|
| author | Heid, Esther |
| author_facet | Heid, Esther |
| contents | <p>Machine learning is increasingly transforming our ability to predict, analyze, and generate chemical reactions, yet the development of robust reaction-centered models remains substantially more complex than molecular property prediction. This talk will introduce a set of recent advances aimed at addressing these challenges and moving toward a general reaction presentation. First, I will introduce ChemTorch [1], a general framework for reaction machine learning that unifies graph neural networks on reaction graphs, neural networks on reaction fingerprint, 3-dimensional graph convolutional architectures for coordinate-based representations, and string encoders for SMILES and related formats. I will also discuss a recent extension that augments graph representations with learned encodings of their 3D environments, enabling highly accurate predictions of barrier heights [2]. Building on these foundations, the talk will present GoFlow [3], a flow-matching generative model for transition-state structures, followed by a steering framework for flow matching [4] that allows fine-grained control over the generated geometries, including enforcement of correct transition-state chiralities. Together, these developments illustrate how predictive and generative deep learning can provide deeper insight into chemical reactivity if they are anchored in principled reaction representations and suitable inductive biases.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18375668 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Predictive and Generative Deep Learning Approaches for Chemical Reactions Heid, Esther <p>Machine learning is increasingly transforming our ability to predict, analyze, and generate chemical reactions, yet the development of robust reaction-centered models remains substantially more complex than molecular property prediction. This talk will introduce a set of recent advances aimed at addressing these challenges and moving toward a general reaction presentation. First, I will introduce ChemTorch [1], a general framework for reaction machine learning that unifies graph neural networks on reaction graphs, neural networks on reaction fingerprint, 3-dimensional graph convolutional architectures for coordinate-based representations, and string encoders for SMILES and related formats. I will also discuss a recent extension that augments graph representations with learned encodings of their 3D environments, enabling highly accurate predictions of barrier heights [2]. Building on these foundations, the talk will present GoFlow [3], a flow-matching generative model for transition-state structures, followed by a steering framework for flow matching [4] that allows fine-grained control over the generated geometries, including enforcement of correct transition-state chiralities. Together, these developments illustrate how predictive and generative deep learning can provide deeper insight into chemical reactivity if they are anchored in principled reaction representations and suitable inductive biases.</p> |
| title | Predictive and Generative Deep Learning Approaches for Chemical Reactions |
| url | https://doi.org/10.5281/zenodo.18375668 |