Machine learning, and in particular its modern incarnation of deep learning (DL) [1, 2], has become an indispensable tool in particle physics, a field that routinely handles vast datasets and nonlinear relationships among observables [3-6]. In recent years, advances in DL have expanded the scope of data-driven progress across the energy, intensity, accelerator, and cosmic frontiers [7, 8]. Despite remarkable advancements, most current DL applications in particle physics are supervised, relying either on Monte Carlo (MC) simulations or on labeled experimental data. However, because simulations cannot fully capture the complexity of the real world, a persistent gap between MC and Data leads to training bias. Direct training on real data, in turn, demands extensive human labeling, which is labor-intensive and hard to scale [9]. For this reason, the development of unsupervised DL [10, 11] is integral for particle physics. Unsupervised learning has achieved remarkable success in tasks such as clustering [12, 13], anomaly detection [14, 15], and learning representations [16, 17].