Abstract
This thesis investigates how artificial intelligence can be used to construct interpretable reduced descriptions of active matter and other non-equilibrium many-body systems. Two complementary levels of description are considered. The first treats topological defects in nematic materials as quasiparticle-like objects whose positions, charges, and orientations must be extracted from high-dimensional experimental and simulated textures. The second treats interacting agents as a minimal active-matter-inspired system in which collective transport rules emerge through repeated local interaction.
For defect analysis, this thesis develops the Machine Eye for Defects (MED), a physics-informed machine-learning framework that combines simulation-assisted data generation, director-field reconstruction, object detection, and transformer-based orientation inference. Synthetic training data are constructed from Frank–Oseen theory and hybrid LBM–Beris–Edwards simulations, while experimental images are processed using SAM-based segmentation, PCA- or gradient-based local-orientation extraction, and interpolation in the doubled-angle representation. The resulting director fields are analyzed by a NanoDet-Plus-based Defect Detection Network and a Vision-Transformer-based Orientation Transformer. MED is shown to identify and characterize defects across epithelial-cell monolayers, actin- and microtubule-based active nematics, molecular-simulation snapshots, active-turbulence simulations, and defect textures with giant cores. In active-turbulence simulations with known ground truth, MED detects 120 out of 123 defects, corresponding to 97.56% accuracy, with most orientation errors below 5∘. In epithelial-cell images, comparison with an order-parameter-based benchmark gives 90.74% defect-identification accuracy, with most defect-core deviations below 2 μm. Moreover, although trained only on ±1/2 defects, MED interprets higher-charge singularities as clusters of half-integer defects whose total charge and inferred structural orientation remain physically meaningful.
For collective rule emergence, this thesis develops a hierarchical reinforcement-learning framework. A DQAttention single-agent policy is first trained by behavior cloning from approximately 108 million A∗-generated samples and then refined under a sparse, terminal-dominated reinforcement-learning reward. The trained policy reaches 100.00% success over 8000 randomized single-agent navigation tests and is subsequently frozen as the navigation backbone of a Social Residual Q-Network (SRQN). In two-agent two-way-street environments, SRQN learns only an interaction-dependent residual correction. Across 144 deterministic crossing scenarios, the agents discover multiple successful traffic-like conventions, including aggressive–cooperative role differentiation, priority–yielding behavior, asymmetric lane preference, and handedness-like rules. Post-differentiation retraining further shows that mature social roles can be reconstructed as reciprocal behavioral modes, while shared-parameter and human-prior experiments demonstrate how individual habits can be transformed into more scalable traffic conventions.
Together, these results show that artificial intelligence can serve not merely as a black-box predictor, but as both a physically informed measurement operator for extracting quasiparticles from complex fields and a computational laboratory for discovering emergent rules in interacting active systems.