Deep Operator Networks (DeepONet) for Reactive Flow

Accurate computational modeling and predictions of reactive flows with stiff chemical kinetics requires computational fluid dynamic (CFD) codes either: 1) with very small timesteps to accurately capture the non-linear kinetics, or 2) with simplified skeletal kinetics that can improve runtime but reduce accuracy. This is a severe limitation that makes accurate modeling of the chemical kinetics very challenging. To accurately model these types of reaction mechanisms in a CFD simulation without a loss in fidelity or accuracy, a data-driven machine learning approach is needed that can accurately predict the chemical kinetics with a large CFD timestep without compromising accuracy. Deep Operator Networks (DeepONet) is an exciting data-driven machine learning technique that can learn nonlinear operators from training data. In the proposed effort, a DeepONet software library will be developed that can be integrated into any CFD code using a flexible software interface. A pre-trained DeepONet database for various reaction mechanisms can then be used to enable CFD codes to accurately simulate reactive flows with stiff kinetics with speeds greater than x1000.