Differentiable GPU-accelerated MRI Simulations

Advanced magnetic resonance imaging techniques, such as RF pulse optimization, increasingly rely on complex, high-performance physical simulations like those defined in KomaMRI.jl. Currently, integrating these simulations into optimization loops is bottlenecked by the lack of native gradient propagation. This project aims to unlock gradient-based optimization within the Koma ecosystem by enabling reverse-mode automatic differentiation (AD) as a first-class feature. By systematically integrating state-of-the-art AD tools, specifically Enzyme.jl and Reactant.jl, this project will make both the standard BlochSimple method and the highly optimized, custom Bloch GPU kernels fully differentiable. Core deliverables include establishing AD correctness baselines against FiniteDifferences.jl, designing a memory-efficient Simulation data structure to minimize device-host transfers, and resolving low-level upstream compatibility issues with CUDA.jl and Enzyme. Ultimately, this project will provide the Julia MRI community with a robust, fully differentiable simulation stack, culminating in an interactive tutorial for 1D RF pulse optimization.