
Project Description
Deep learning can reconstruct daily river flow from climate data alone, filling records where gauges have stopped. Such models learn the climate–discharge relationship from past decades, then apply it forward in time. But catchments do not stand still: glaciers retreat, snow declines, land cover changes. A model that never observed those changes can keep producing a plausible, well-shaped hydrograph while missing the underlying trend, a failure conventional performance scores do not reveal.
This project will explore when such models remain valid. Using long-record catchments and climate reanalysis held by the supervisory team, the student will run differential split-sample experiments: train on early periods, then test whether the model recovers multi-decadal trends rather than only seasonal shape. Methods include out-of-distribution detection on the neural network’s internal states, flagging when inputs fall outside the range it was fitted on; catchment descriptors that evolve through time rather than being held fixed; and differentiable hybrid models, in which a physically constrained storage module is trained end-to-end with a neural network so that mass is conserved and storage depletion is representable.
Reconstructed flow records now inform water allocation, infrastructure design and adaptation planning. This project supplies the credibility layer such products lack, and a transferable method for testing whether any environmental AI model still holds as its system changes.
Project Specific Training
Training is delivered principally one-to-one by the supervisory team, who provide the existing code base, data pipelines and trained baseline models at the outset. The student receives structured instruction in deep learning with TensorFlow or PyTorch, hybrid physics–machine learning modelling, hydrological model evaluation and reproducible scientific computing, reinforced through weekly meetings. High-performance computing training is delivered by the University’s research computing service. The student will attend one external short course or summer school on machine learning for environmental science, and will present at an international conference. Group code reviews with other students on related projects provide additional peer training.
Potential Career Trajectory
The project builds a combination of skills that is currently scarce: rigorous machine learning practice grounded in a physical environmental engineering, together with a specialism in model reliability and validation. Within academia it is direct preparation for postdoctoral research in hydrology, machine learning for Earth systems, and climate impact modelling. Outside academia, graduates are well placed for national hydrological and meteorological services, environmental regulators and water utilities; engineering and environmental consultancies delivering flood risk and water resource assessment
Access and Inclusion Statement
This is a fully computational project. All data are existing digital archives already held by the supervisory team, so there is no fieldwork, no field sampling, and no overseas travel requirement at any stage. Analysis runs on remotely accessible computing facilities, so the student is not tied to a fixed laboratory. No prior programming experience is assumed, although knowing Python will be an advantage, but it is not necessary. Training is structured and paced to the individual, and the student begins from a working code base rather than a blank page, so applicants from quantitative backgrounds outside hydrology are not disadvantaged. The supervisory team will work with the student and with disability services to configure the development environment, and code review can be conducted in writing or verbally according to preference.
Project supervisor/s
Ashraf Ahmed
Civil and Environmental Engineering
Brunel University of London
ashraf.ahmed@brunel.ac.uk
https://www.brunel.ac.uk/people/ashraf-ahmed/research
Muhammad Shafique
Civil and Environmental Engineering
Brunel University of London
Muhammad.Shafique@brunel.ac.uk
