PhD Studentship: Physics Informed Machine Learning for Climate Impacts on Hydrology

Hydrological modelling has a long legacy of development and as such there are a huge range of models to choose from for practical use. One of the key differences between hydrological models is the level of physical representation of catchment processes within the model code. Machine learning (ML) models are entirely data-driven and contain no pre-conceived representation of catchment processes. Conceptual models represent processes in a simplified way that are parameterised and calibrated to observational data. Physically based (PB) models codify known physical laws into a single modelling framework. Each model structure has strengths and weaknesses. Recent studies have highlighted the superior performance of Machine Learning (ML) models over conceptual and PB models in replicating historical river flows, indicating their potential for more effective operational use, yet water managers remain wary of ML models due to their opaque nature, raising concerns about process visibility. Conversely, PB models offer explicit representation of known physical processes, and have been shown to simulate more robust projections of future river flows under climate change. Current operational methodologies predominantly rely on conceptual models, lacking the sophistication of more advanced techniques.

Emerging research suggests that hybrid ML and PB models hold promise for achieving even better historical simulations, with the added bonus of improved process understanding and robustness for use in climate impact studies. However, this area remains largely unexplored in the hydrological domain. This PhD opportunity therefore aims to explore the area of PB+ML hydrological modelling in the UK context and address several key research objectives:

• Investigate the reasons behind the superior performance of ML models over physically based models and how they leverage data more effectively.
• Explore whether ML models emulate physical processes absent in physically based models and if ML models can identify expressions of these processes.
• Assess the applicability of existing Physics-Informed Neural Network (PINN) and hybrid ML+PB models in hydrological contexts, determining the most appropriate framework.
• Evaluate whether a PB+ML model outperforms existing national models in the UK.
• Develop a national-scale PB+ML model for the UK and assess its confidence in predicting flows in a variety of catchments.
• Examine the ability of a PB+ML model to robustly project future changes in floods and droughts.
• Embark on this exciting journey to push the boundaries of hydrological modelling and contribute to solutions for pressing water management challenges. Apply now to be at the forefront of cutting-edge research in the field.

Eligibility

Applicants should have, or expect to achieve, an excellent academic record (UK First-class or 2.1 honours or international equivalent depending on the funding source) in Engineering, Earth Sciences, Computing or another related physical science discipline (MSc, MSci or BSc). You should have appropriate experience in hydrology, modelling or machine learning and an interest in developing your modelling skills. Some knowledge or previous experience in flood and drought management or computational modelling would be helpful, but is not an essential since you will receive training in all the relevant techniques. You will be encouraged to attend national and international conferences to share your research.

Advert information