Speaker
Description
The design of modern high-energy physics detectors is a highly intricate problem: maximize the physics performance while balancing various manufacturing constraints. A holistic solution consists in translating the design process into an optimization task suitable for Machine Learning methods. The main difficulty however lies in producing meaningful gradients from the detector simulation.
The AIDO framework solves this issue by training a diffusion-based surrogate model on various detector geometries and interpolating the expected performance across different configurations. This enables gradient descent in the surrogate parameter space for both continuous and discrete parameters such as materials and layer numbers.
As a demonstration, we generate an optimal sampling calorimeter by maximizing its energy resolution starting from a random initial composition. The model was trained with an automatic scheduling tool on hardware partially provided by KCETA.