A PhD student is building artificial intelligence models that learn the physical laws of blood flow and heart tissue mechanics, rather than just learning from data alone. This matters because current computer models of the heart are too slow and computationally expensive for real-time clinical use. Doctors need rapid predictions—for example, to assess how a patient’s heart will respond to a treatment during a consultation—but the underlying equations of fluid dynamics and soft tissue mechanics take hours or days to solve with traditional numerical methods. The student aims to create surrogate models that approximate those solutions in seconds, while also requiring less training data and producing results that doctors can interpret because the models obey known physics. If successful, this could speed up cardiovascular diagnostics and treatment planning. The student will also generate synthetic heart images using generative AI—effectively creating realistic ‘deepfake’ left ventricles—to supplement real cardiac MRI scans. This larger, augmented dataset should improve the accuracy of the emulators. The work is primarily methodological, advancing fundamental techniques in physics-informed machine learning and uncertainty quantification, but it has a clear path toward practical impact in clinical settings where rapid, reliable predictions are currently out of reach.
View original technical description
The objective of the PhD project is parameter estimation, sensitivity analysis and uncertainty quantification in complex cardio-physiological models. To this end, the PhD student will be developing advanced surrogate models to emulate the underlying fluid dynamics or soft tissue mechanics physical models. Rather than building these surrogate models in a purely data-driven manner with established statistical or machine learning methods, the student will be integrating the physical laws themselves into the surrogate model building process, to achieve the following three objectives. 1) Higher data efficiency: by using physical laws to constrain the solution space, the student will aim to achieve good performance with less training data. 2) Interpretability: by incorporating known physical principles, the student will aim to build more interpretable machine learning models, as their predictions will adhere to the mathematical structure of physical laws. 3) Reduced computational costs: the student's aim will be to approximate solutions to complex partial differential equations without traditional numerical solvers, which is valuable in clinical real-time applications where rapid predictions are needed. The student will further apply generative AI techniques to produce 'deepfake' images of the heart, which will be used to supplement the existing set of real Cardiovascular Magnetic Resonance (CMR) scans of the left ventricle of the heart. Using an existing database of CMR images, the student will use generative adversarial networks to generate deep fake left ventricles, with the ultimate objective to produce more synthetic left ventricle geometries and thereby boost the training set size and emulator accuracy. To adequately cover the broad range of methodologies, the student will be supervised by a team of four supervisors, with two supervisors providing support on the physics-informed machine learning part of the project, and the other two supervisors leading the work on generative AI techniques.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know