Uncovering Underlying Dynamics in Non-Equilibrium Systems Through Complementary Modeling Perspectives

Biological active matter self-assembles and operates far from equilibrium, and its behavior spans many scales. No single modeling perspective captures both the fine-grained behavior of a specific non-equilibrium system and the general principles that govern it. A combination of complementary perspectives, from different regions of the space of modeling approaches, can recover both. This thesis develops that idea across three such systems, using minimal thermodynamic models, dimensionality reduction, and generative machine learning. Together, these approaches characterize behavior in parameter regimes for which detailed data is not available. A minimal thermodynamic model of branched actin self-assembly captures the growth velocity under load, and its entropy production bounds the structures the network can access. The bound reveals a dissipation gap. This energy cost, beyond the work of moving the load, maintains the free-end fluctuations the network uses to adapt. For a chemical reaction-diffusion system producing traveling waves, a learned latent space both predicts the phase boundaries between wave behaviors from a small number of simulations and reduces the dynamics of seven chemical species to a single collective variable. For a simulated actomyosin cortex, a generative diffusion model trained on coarse-grained static images at a few filament turnover rates infers a non-monotonic curvature trend at rates it never saw. A minimal motor-binding model traces the nonlinearity to a diffusion-limited shift in motor binding rates. These perspectives range from data-driven inference to interpretable mechanism, trading fine-grained detail against generality in different ways. The same active matter can be treated as a region to bound, a trend to predict, or a distribution to sample.