Context:
Safe and reliable operations of future fusion power plants require rigorous qualification of materials & components. But the current approaches to qualification for nuclear systems is a multi-million pounds worth decades-long process because it is expected to experimentally quantify all the expected failure modes under fusion relevant conditions for safety/reliability. The process is, therefore, largely limited by lengthy timescales to perform long-lead testing such as multi-year long thermal creep tests and neutron irradiation campaigns using test reactors. These are needed to satisfy requirements imposed by engineering codes/standards being developed for fusion with the American Society for Mechanical Engineers (ASME) or through the French approach using the RCC-MRx code. If qualification process takes several years, which currently does, then this is a major bottleneck to the vision of accelerated fusion deployment globally. In this context, this EngD will explore the role of AI/ML methods to develop material & component-scale engineering constitutive models using existing degradation data in literature and help pave a pathway for an accelerated qualification framework.
Current state-of-the-art:
A key failure mode from which fusion first-wall/blanket structures need to be protected against is low-temperature hardening-embrittlement (LTHE) (Bhattacharya et al., 2022), where materials & components show increase in yield stress with near-total loss of ductility and fracture toughness. This is a major design limiting challenge for all fusion systems, but especially for water-cooled DEMO-like designs (Marian, Fitzgerald & Po, 2020) – that has warranted over 40 years of R&D collecting LTHE data on candidate materials like reduced activation ferritic martensitic (RAFM) steels. Microstructural origins of LTHE are postulated to be radiation-induced defects causing highly localised and heterogeneous plastic deformation (Marian, Fitzgerald & Po, 2020). Currently, no proper engineering models exist that can guide designers to prevent against LTHE failure. A few models using classical approaches have been historically postulated that rely upon “best fits” through experimental neutron irradiation data – but they are often not multi-scale physics-informed and may lack the sophistication to incorporate all the key variables of complex experiments, such as neutron irradiation experiments in materials test reactors (MTRS).
Project:
Macroscopic understanding of the physical and chemical processes can be used to support ML models, for instance through physics-informed neural networks (PINNs) for solving nonlinear equations (Raissi et al., 2019) or through the optimisation of process parameters in physics-informed reinforcement learning (Mohamed et al., 2025). The aim of this EngD programme is twofold. First, build physics-informed models that are able to capture the general properties of the materials & bulk-scale components in relevant geometries under investigation – such as a fusion breeder blanket. Second, to use the information at microscopic level (structural properties) and generalise the high-level behaviour – with a focus towards AI informed constitutive model and develop structural design criteria to prevent against failure due to LTHE. In particular, this study will explore how to detect and classify the defects and associate various deviations with changed material properties. This EngD programme will approach the problem as anomaly detection where, among other physics-informed ML, it will build a generative model of defect-free components and detect defects as outliers from the normal distribution. Specifically, for density estimation possible frameworks include generative models, e.g., variational Autoencoders (VAEs), generative adversarial networks (GANs), latent diffusion models, which will be trained to learn the probability distribution of the normal data. For anomaly scoring, once trained, these models will be used to estimate the likelihood or probability of new data points. Stitching the actual experimental data, merged with AI-informed synthetic data will then generate the relationships between the different engineering parameters for mitigating failure – such as linking defect properties to stress-strain relationships.
We are seeking a highly motivated candidate with a background in Computer Science, Automatic Control or related disciplines. Essential: strong mathematical foundations, and strong programming skills. Desirable: ability to work in an interdisciplinary topic, understanding of dynamical systems.
Collaborations:
This EngD is a collaboration between University of Birmingham, University of California Los Angeles, University of Manchester with industry partners: Electric Power Research Institute & Tokamak Energy.
For informal queries, reach out to Dr. Leonardo Stella (l.stella@bham.ac.uk), Prof. Arun Bhattacharya (a.bhattacharya.1@bham.ac.uk) and Prof. Ales Leonardis (a.leonardis@bham.ac.uk).
This EngD project is funded by the Fusion Engineering CDT and hence the student will be based at The University of Birmingham, but should expect to engage fully with the 3-month training programme within the Fusion Engineering CDT at the start of the course. CDT training will be delivered across the CDT partner universities at Sheffield, Manchester, Birmingham and Liverpool. For further information about the CDT programme, please visit the website or send an email to hello@fusion-engineering-cdt.ac.uk.
Apply for this project at The University of Birmingham here.
fusion_cdt
