Scientific Machine Learning

This dissertation topic is available for Czech studies only.

For decades, the behavior of systems in the physical world has been modeled by numerical models based on the vast scientific knowledge about the underlying natural laws. However, the increasing capabilities of machine learning algorithms are starting to disrupt this landscape. With large enough datasets, they can learn the recurring patterns in the data.

However, a pure machine learning model usually has poor interpretability, needs a lot of data to train which can be hard to come by in many scientific domains, and might not be able to generalize properly. These concerns are being addressed by the field of scientific machine learning (SciML) – an emerging discipline within the data science community. It introduces scientific domain knowledge into the learning process. SciML aims to develop new methods for scalable, robust, interpretable and reliable learning.

Physics-informed neural networks are a part of SciML - we include the physical constraints in the model through appropriate loss functions or tailored interventions into the model architecture. Through physics-informed machine learning, we can create neural network models that are physically consistent, data efficient, and trustworthy.

The goal of the research is to explore how to incorporate scientific knowledge into the machine learning models, thus creating hybrid models based on SciML principles that include both data-driven and domain-aware components. The research could also be directed towards a combination of SciML and transfer learning (that reuses a pre-trained model on a new problem). The aim of such a combination is to take advantage of both approaches.

SciML can be applied in many domains – we focus mainly on power engineering, e.g. supporting the adoption of renewables and on Earth science with emphasis on positive environmental impact improving climate resilience, but any other domain could be selected.

Relevant publications:

• Kloska, M., Rozinajova, V., Grmanova, G. Expert Enhanced Dynamic Time Warping Based Anomaly Detection. Expert Systems with Applications (2023) https://arxiv.org/pdf/2310.02280.pdf
• Pavlik, P., Rozinajova, V., Bou Ezzeddine, A. Radar-Based Volumetric Precipitation Nowcasting: A 3D Convolutional Neural Network with UNet Architecture. Workshop on Complex Data Challenges in Earth Observation 2022 at CAI-ECAI (2022) https://ceur-ws.org/Vol-3207/paper10.pdf

The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.