Scientific Machine Learning and AI
About
We view "scientific machine learning and AI" as an interdisciplinary endeavor that combines scientific computing and machine learning to solve domain-specific problems in the physical sciences and engineering. Our group focuses on developing methods for physics-informed machine learning, hybrid and surrogate modeling, purely data-driven (black-box) modeling, and foundation models for multimodal scientific data. We aim to create and train ML and AI models tailored to specific classes of scientific problems, considering factors such as required accuracy and available data.
Research Topics
Our major research areas include:
1. Hybrid and Surrogate Models
Surrogate (or forward) models utilize simulation datasets or microscopy images to learn complex spatio-temporal relationships to scalar or tensorial values or fields. For example, predicting a 2D field of pressure values in a fluid where the Reynolds number is desired, or determining accumulated plastic strain from a grain microstructure. Our goal is to develop computationally efficient surrogate models that meet the accuracy requirements of scientific problems.
2. Inverse Modeling
We address inverse problems, which may lack unique solutions but can be transformed into well-posed problems through deep learning-based regularization and the incorporation of physics constraints. There, analyzing and tuning the latent space representation is one of the used approaches.
3. Foundation Models for Multimodal Scientific Data
In the context of photovoltaics, we are part of a larger team who is developing a foundation models for multimodal scientific data. This involves creating encoders that effectively bind different data modalities - such as spectral data, imaging, and numerical simulations -- to enhance the analysis and understanding of photovoltaic materials and systems.
Across all our projects, we focus on creating ML and AI models that are not only accurate and computationally efficient but also adaptable to a wide range of scientific problems and that ultimately help to solve concrete scientific problems.
