Descriptive and Predictive Modeling of Chaos in Cold Atom Physics using Explainable Deep Learning

Ultracold atom clouds display intricate dynamics that can appear chaotic and lack a comprehensive explanatory framework. This dissertation investigates photon bubble turbulence in a 85Rb atom cloud confined in a magneto-optical trap, using recent advances in artificial intelligence to move beyond the limitations of classical modeling. Reduced-order methods and sparse regression provide interpretability but fail to capture long-term dynamics or generalize across regimes.

To overcome these issues, deep learning models were developed, with a focus on convolutional autoencoders for reconstructing and analyzing atom cloud images. Results showed that these architectures consistently outperformed classical baselines, even in lowdimensional latent spaces, demonstrating the ability of convolutional representations to extract physically meaningful features. Latent space analysis revealed that the apparently chaotic turbulence is underpinned by compact structures, suggesting a simple order to chaos.

Classification experiments showed that detuning values, determined by laser frequencies in the trap, can be reliably inferred from images, exposing discriminative features tied to light–atom interactions. Explainability was introduced through Grad-CAM, linking classification decisions to density structures in the clouds. The findings confirmed two dynamical regimes, stable and turbulent, separated by a phase transition, with further evidence of a transitional regime near resonance. Overall, explainable deep learning emerges as a powerful framework for predictive and descriptive modeling of ultracold atom systems, while also deepening our understanding of turbulence, phase transitions, and light–atom coupling.