Abstract: Bridging Physics and AI - Learning Pedestrian Dynamics from Video Data. Deep neural networks (DNNs) have rapidly transformed many fields, often replacing traditional explicit models despite their “black box” nature. However, explicit models encapsulate valuable human knowledge, and rather than discarding them, a powerful research trend has emerged: combining them with DNNs. Initially applied to fluid dynamics in 2016, this approach—now known as AI for Science—has since expanded across disciplines. In this talk, I will demonstrate how physics-informed deep learning extends far beyond solving physics problems. I will showcase how we leverage this approach to analyse pedestrian dynamics from video data, tackling some of the most challenging real-world scenarios, including ultra-dense crowds where the risk of crushes is high. By fusing physics with AI, we uncover deeper insights and push the boundaries of what’s possible in modelling complex human behaviours.