idw – Informationsdienst Wissenschaft
Nachrichten, Termine, Experten
Nachrichten, Termine, Experten
BIFOLD funds agility projects to support flexible, short-term research collaborations and academic networking. The project “Guided Exploration of Chemical Space with Deep Neural Networks and Bayesian Optimization” recently concluded, advancing material discovery by integrating deep neural networks with Bayesian optimization to improve energy efficiency and reliably identify stable crystal structures in vast chemical spaces. The project was jointly conducted by BIFOLD researchers Dr. Kristof Schütt and Stefaan Hessmann together with their Japanese partners Tomoki Yamashita (Nagaoka University of Technology) and Tamio Oguchi (Osaka University, NIMS).
Guided exploration of chemical space with DNN and Bayesian Optimization
Global energy optimization is key challenge in this process
The development of new, stable materials has enabled significant advancements across various research fields, including the discovery of solar cells, catalysts, superconductors, hardware components, and batteries. A key challenge in this process is global energy optimization, which is essential for identifying stable crystal structures in the vast space of possible atomic configurations. The Agility project Guided Exploration of Chemical Space with Deep Neural Networks and Bayesian Optimization has contributed to the advancement of current research in this area. The project was jointly conducted by BIFOLD researchers Dr. Kristof Schütt and Stefaan Hessmann together with their Japanese partners Tomoki Yamashita (Nagaoka University of Technology) and Tamio Oguchi (Osaka University, NIMS).
The scientists developed a fully automated method for active learning with deep neural networks that can identify local (and global) energy minima of crystal structures from scratch. The primary focus of the project was to utilize active learning MLFF ensembles (Machine Learning Force Field Ensembles: a method that allows quantification of uncertainties in the predictions) to speed up structure relaxation of randomly generated candidate structures towards their local minima. Here the uncertainty estimates prevent faulty predictions by the DNNs and allows the selection of the most important training data.
This enables a fully automated workflow that does not require any prior knowledge of the target system. Ultimately, this allows fast and accurate detection of the energy minima of crystal structures, while significantly reducing the number of expensive calculations. The method can also be efficiently parallelized, making it ideal for high-performance computers. The project therefore contributes to the development of efficient methods for crystal structure searches.
Software Development and Molecular Generation
Alongside this primary research effort, the project has collaborated with three related projects in molecular generation and software development:
*Open-Source Software Development (SchNetPack 2.0): This open-source package provides a framework for developing deep neural networks tailored to chemical calculations. It enables researchers to train and apply neural networks for both molecular and material simulations.
*Autoregressive Molecular Generation: This initiative focuses on using autoregressive models to generate three-dimensional molecular structures with specific chemical and structural properties. These approaches allow for the targeted design of molecules with predefined physicochemical characteristics, with potential future applications in crystalline structures.
*Diffusion Models for Molecular Relaxation (MoreRed - Molecular Relaxation by Reverse Diffusion): This project presents an innovative statistical approach to molecular relaxation by treating unstable structures as noisy versions of their stable counterparts. By leveraging generative diffusion models, the method significantly reduces computational effort while improving the efficiency of structural optimization.
The Agility project achieved to integrate active learning of DNNs for crystal structure algorithms, where machine learning force field ensembles accelerate structure relaxation and energy optimization. SchNetPack 2.0 serves as a versatile open-source platform, offering a comprehensive suite of tools for machine learning applications in materials and molecular sciences. Additionally, molecular generation algorithms, including autoregressive and diffusion models, introduce new possibilities for designing molecules with specific properties that could be extended to material structures. These combined approaches enhance computational techniques for material discovery and accelerate the development of novel materials with tailored properties.
Stefaan Hessmann
BIFOLD / TU Berlin Machine Learning Group
stefaan.hessman@tu-berlin.de
https://web.ml.tu-berlin.de/author/stefaan-hessmann/
S.S.P. Hessmann, K.T. Schütt, N.W.A. Gebauer, M. Gastegger, T. Oguchi, and T. Yamashita. "Accelerating crystal structure search through active learning with neural networks for rapid relaxations." npj Computational Materials (accepted, 2025). arXiv:2408.04073.
K.T. Schütt, S.S.P. Hessmann, N.W.A. Gebauer, J. Lederer, and M. Gastegger. "SchNetPack 2.0: A neural network toolbox for atomistic machine learning." The Journal of Chemical Physics, 158(14) (2023).
N.W.A. Gebauer, M. Gastegger, S.S.P. Hessmann, K.-R. Müller, and K.T. Schütt. "Inverse design of 3D molecular structures with conditional generative neural networks." Nature Communications, 13(1), 973 (2022).
K. Kahouli, S.S.P. Hessmann, K.-R. Müller, S. Nakajima, S. Gugler, and N.W.A. Gebauer. "Molecular relaxation by reverse diffusion with time step prediction." Machine Learning: Science and Technology, 5(3), 035038 (2024).
3D renders of the lowest energy minima of Al₁₆O₂₄ crystal structures, identified using the newly dev ...
BIFOLD
Merkmale dieser Pressemitteilung:
Journalisten, Wissenschaftler
Chemie, Informationstechnik
überregional
Forschungsergebnisse, Forschungsprojekte
Englisch
3D renders of the lowest energy minima of Al₁₆O₂₄ crystal structures, identified using the newly dev ...
BIFOLD
Sie können Suchbegriffe mit und, oder und / oder nicht verknüpfen, z. B. Philo nicht logie.
Verknüpfungen können Sie mit Klammern voneinander trennen, z. B. (Philo nicht logie) oder (Psycho und logie).
Zusammenhängende Worte werden als Wortgruppe gesucht, wenn Sie sie in Anführungsstriche setzen, z. B. „Bundesrepublik Deutschland“.
Die Erweiterte Suche können Sie auch nutzen, ohne Suchbegriffe einzugeben. Sie orientiert sich dann an den Kriterien, die Sie ausgewählt haben (z. B. nach dem Land oder dem Sachgebiet).
Haben Sie in einer Kategorie kein Kriterium ausgewählt, wird die gesamte Kategorie durchsucht (z.B. alle Sachgebiete oder alle Länder).
