This work reports an application of physics-informed machine learning models on reconstructing key parameters of acoustically forced, time-varying laminar sooting flames, highlighting the potential of the machine learning methods as a complementary tool to conventional laser diagnostics. First, a physics-informed neural networks (PINNs) model was developed to reconstruct the fields of velocity and temperature in the region where is inaccessible with laser-based diagnosing methods due to soot scattering. The PINNs model was trained using experimental data from planar laser diagnostics and constrained with the momentum and energy conservations. The model shows effective capability of fulfilling the velocity and temperature fields. Second, an Autoencoder (AE)-based Deep Operator Network (DeepONet), also as a physics-informed model, was developed to predict the planar distribution of soot volume fraction in the flames. The AE-DeepONet framework was trained using planar images of temperature and hydroxyl radical (OH) with a hybrid way by combining physics-informed and data-driven approaches. The AE-DeepONet model outperforms the conventional data-driven-only machine learning models. The results show that, constrained by physical laws, machine learning based models can properly predict soot distribution, velocity and temperature in unsteady laminar flames, shedding light on the physics-informed machine learning methods as a complement to laser diagnostics.
Read full abstract