Research topics

Diffraction methods are widely used to determine the structures of solid materials, small molecules, and biological samples. In recent years, electron diffraction (ED) has emerged as a promising alternative to traditional approaches (such as X-ray and neutron diffraction). One major benefit is that ED requires much smaller crystals, and data can be collected using standard transmission electron microscopes. However, analyzing ED data can be challenging because electrons interact strongly with matter (often referred to as dynamical effects), which can make structure determination less straightforward.

To overcome these challenges, we are using machine learning – including convolutional neural networks (CNNs) and graph neural networks (GNNs) – to automate critical steps in the ED data analysis process. Specifically, we aim to:

• Automatically identify (“index”) diffraction patterns,
• Classify data in real time during collection, and
• Refine structure solutions more accurately.

This project is a collaboration between the research groups of Jun.-Prof. Dr. Julia Westermayr (Leipzig University) and Prof. Dr. Xiaodong Zou (Stockholm University). Data collection is carried out in both Leipzig and Stockholm, taking advantage of specialized instrumentation at each site.

Please take a moment, breathe deeply, and describe what you smell. Not so easy, is it?

Humans excel at distinguishing odors, but we often struggle to put them into words. It is no surprise that we can map the perception of color and pitch to wavelength and frequency, yet the relationship between odor descriptions and the chemical structure of remains a puzzle. The olfactory stimuli- and description-space is complex and high dimensional. Machine learning has demonstrated its capabilities under similar conditions in fields like natural language processing. Thus, we want to explore the potential of machine learning in fragrance research. 

Our mission aims to:

• Language Scent Mapping: Correlating perfumery language and odor perception using natural language processing (e.g. large language models)
• Odor prediction: Using classical and deep machine learning in combination with fingerprints and graph neural networks, to predict the odor of complex fragrance mixtures from the chemical structure
• Fragrance Optimization: Applying our models on fragrance formulas

This work primarily focuses on developing machine learning tools to simulate and design molecular systems in various contexts. Recent efforts have involved integrating excited-state dynamics, incorporating molecular interactions with external fields, and including quantum nuclear effects in machine learning architectures. Specifically, a multipole expansion was integrated into the MACE framework to create FieldMACE, enabling efficient modeling of long-range interactions and environmental effects in large systems while using transfer learning to reduce data requirements. An excited-state variant called X-MACE was also created, employing a DeepSets autoencoder to handle challenging non-smooth regions near conical intersections. Ongoing research explores the use of transfer learning and other frameworks to model quantum nuclear effects in scenarios involving hydrogen isotopes. Additionally, work has been done to combine genetic optimization with generative modeling to refine reaction environments, and reinforcement learning (using an actor-critic approach and models such as PaiNN) has been applied to identify transition states and minimum energy pathways for reactions.

Generative machine learning is changing how we discover new molecules by making exploration of the massive chemical space more efficient. The number of possible compounds is practically infinite, and traditional methods can't keep up with the scale. By learning patterns from existing data, generative models can suggest new molecules with optimized properties, cutting down on trial-and-error. In drug design, they help find bioactive compounds with better selectivity and fewer side effects. In catalysis, they can propose materials that make reactions more efficient, saving energy and reducing costs. 

Our project applies this concept to designing optimized reaction environments that enhance chemical transformations. Instead of searching for single catalyst molecules, we focus on engineering entire molecular surroundings that stabilize key reaction intermediates and lower activation barriers. To achieve this, we integrate generative machine learning with a genetic optimization framework, iteratively refining reaction environments to maximize catalytic efficiency. A key component of our approach is the use of diffusion-based generative models, which allow for flexible and unbiased exploration of molecular structures by gradually refining random noise into chemically meaningful configurations. We combine this with MACE, an equivariant machine learning force field, to predict interaction energies and guide the selection of promising molecular environments.