Accelerating burn region modeling in reactive flows using latent space dynamics modeling

Modeling the spatiotemporal dynamics of reactive flows is challenging due to their inherently multi-scale, nonlinear, and shock-driven behavior. While high-fidelity simulations provide accurate insights, their computational cost renders them impractical for parameter space exploration tasks, such as determining optimal initial burn configurations for effective fuel ignition and sustained reaction propagation. Recent advances in data-driven modeling and model reduction offer a path toward efficient surrogates, but their application to reactive, shock-dominated flows remains limited.

In this talk, we present a latent space dynamics framework for modeling reacting flows and assessing ignition success across varying initial conditions. The proposed approach consists of two key components: (1) spatial dimensionality reduction using implicit neural representations informed by sparse sensor data, and (2) temporal modeling of the low-dimensional latent dynamics via neural ordinary differential equations (neural ODEs). This formulation enables fast, yet expressive modeling of the reactive flow behavior, even from spatiotemporally sparse inputs.

We demonstrate the effectiveness of the method using sparse data from flow simulations and quantify cost savings. The learned latent dynamics accurately predict key features of reaction propagation and enable efficient exploration of system parameters critical to ignition success. Additionally, we show that incorporating physics-based constraints, such as conservation principles, enhances model robustness when data is limited. Overall, this work highlights the promise of latent space dynamics models as powerful and interpretable surrogates for reactive flow simulations, with applications in rapid diagnostics, control, and design of systems.