Majid Rahmanov
I believe I deserve the KIVI EngD Award because I translated a demanding fusion diagnostic
problem into a working software prototype, connecting machine learning with the constraints of
experimental hardware. At DIFFER, in collaboration with Eindhoven University of Technology, I
took responsibility for the engineering path from understanding stakeholder needs to
implementing and evaluating GPU reconstruction.
The challenge was to reconstruct where light originates in the plasma exhaust of MAST-U quickly
enough to support future control. The project's existing iterative methods took approximately 50 to
80 milliseconds per frame, against a three-millisecond reconstruction budget. I combined
calibrated camera geometry with a U-Net and trained it on synthetic emission patterns, including
a check that reconstructed maps reproduce the camera measurements. My contribution connects
this modelling approach to data generation, model export, benchmarking and native C++
integration.
The results demonstrate both technical progress and engineering judgement. The best evaluated
model achieved structural similarity of 0.930 on held-out synthetic data. A smaller development
configuration averaged 2.31 milliseconds per frame. The full-camera A100 C++ prototype
averaged 5.82 milliseconds across six frames, establishing native execution while identifying the
remaining optimisation gap. I documented these distinct results and their limits so that
subsequent development can target the actual bottlenecks. Broader experimental validation and
sustained performance testing remain necessary before use in live control.
I also took ownership of decisions that connected disciplines. When the initial sparse
backprojection implementation could not be exported through the selected toolchain, I adapted
the architecture to a GPU-resident dense operation and evaluated the resulting memory and
performance trade-offs. I worked with DIFFER scientists and engineers and my TU/e supervisor
to align reconstruction quality, runtime interfaces and hardware constraints. I incorporated
scientific feedback from UKAEA and EPFL collaborators. This required explaining design choices
across specialties and maintaining a shared understanding of what the evidence established.
The immediate value is a modular, documented reconstruction pipeline that DIFFER can extend
without rebuilding the entire development process. Its longer-term potential is to support faster
plasma exhaust diagnostics and research into protecting components in future fusion devices.
The work contributes a concrete engineering step towards that goal, with a clear route for further
validation.
This project reflects the contribution I want to make as an engineer: turning advanced
computational methods into useful scientific tools, taking responsibility for implementation and
giving collaborators evidence they can build on. It brings together practical innovation, a reusable
result for fusion research and collaboration across physics, software engineering and highperformance computing

