Abstract: HEP data analysis depends on high-fidelity detector simulators used to translate sensor-level information into physics variables of interest. In this seminar, I will discuss how differentiable simulations allow us to rethink traditional HEP workflows by making simulation parameters directly tuneable through gradient-based optimisation. I will introduce LUCiD, an end-to-end differentiable simulation framework for optical particle detectors. LUCiD matches the performance of established non-differentiable software for Super-Kamiokande while substantially simplifying traditional workflows and enabling new analysis strategies. Moreover, the principles behind LUCiD can be readily adapted to both Cherenkov and scintillation detectors across a wide range of scales, allowing the same framework to model experiments ranging from WCTE, IWCD, Super-Kamiokande, and Hyper-Kamiokande to JUNO and neutrino telescopes such as KM3NeT and IceCube.
The continued development of simulation and inference techniques also depends critically on well-documented public datasets that enable method development, reproducibility, and meaningful comparisons through common benchmarks. As an open-source framework, LUCiD can generate realistic simulated events across different detector technologies and experimental configurations. In the seminar, I will also discuss ongoing efforts to produce large-scale public datasets spanning unprecedented detector diversity and label richness, with the goal of fostering community-wide advances in detector calibration and event reconstruction.