KIPAC Seminar: From Underground Detectors to Astrophysical Transients: A Rare Journey Through Physics and Machine Learning

Many of the most important questions in modern physics—from dark matter searches and neutrinoless double-beta decay to time-domain astronomy—depend on detecting extremely rare signals buried in overwhelming backgrounds. Addressing these problems relies on large-scale Monte Carlo simulations, yet achieving the required statistical precision rapidly becomes computationally prohibitive. As a result, simulation itself is increasingly a limiting factor in experimental design, data analysis, and scientific interpretation. At the same time, standard machine-learning methods offer limited relief in this regime. When signals are rare but not geometrically distinct from background, conventional representation learning and anomaly detection approaches tend to collapse toward dominant patterns, leaving subtle rare-event signatures statistically invisible. In this talk, I will discuss recent work on preserving minority structure and enabling forward–inverse modeling. More broadly, these challenges reflect a growing need across fundamental physics for adaptive simulation controllers in which physics-based simulators and learned surrogates operate together, dynamically allocating computational effort based on uncertainty, verification, and required levels of physical accuracy.