Designing high-energy physics experiments is a high-dimensional, constrained optimization problem that is still largely tackled through long iterative cycles of simulation, reconstruction, and expert-driven trial and error. This workflow makes global optimization difficult: choices are often tuned locally for individual sub-detectors, the feasible design space can be vast and discontinuous, and performance landscapes can contain multiple local minima and disconnected optima. In this talk, I will frame detector design explicitly as an optimization problem and review the growing interest in AI-driven approaches, with a comparative view of different methodological families. As a concrete case study, I will use the SHiP active muon shield, then generalize the discussion to other sub-detector design tasks. I will focus in particular on reinforcement learning as a flexible formulation for mixed discrete and continuous decisions with sequential and combinatorial structure, where constraints and competing objectives can be handled naturally. Finally, I will discuss current limitations, especially the role of uncertainties: stochastic detector response, systematic shifts from physics modeling and reconstruction, and domain mismatch between simulation and reality. I will outline how uncertainty-aware objectives and robust or probabilistic optimization strategies could be incorporated into the design loop.
