A physics-informed neural network for aiding the acquisition of high-fidelity multiphysics fields in gas-phase combustion reacting flows without pre-training

Acquiring high-fidelity multiphysics data in combustion reacting flows remains challenging due to synchronization difficulty, high cost, and significant uncertainty inherent in non-intrusive laser diagnostics. To alleviate these challenges, we propose CRF-PINN (combustion reacting flow physics-informed neural network), a framework that integrates fundamental physical principles, including the Navier–Stokes equations, multi-species transport equations, and global chemical kinetics, with low-cost experimental data (e.g., single-field, low-resolution, or noisy measurements). CRF-PINN eliminates the need for collecting training datasets and pre-training by enforcing the governing equations as soft constraints within our newly developed multi-receptive-field convolutional neural network. We demonstrate its capabilities in three key scenarios: (1) cross-field translation (e.g., inferring velocity from temperature/species, temperature from velocity/species, and species from velocity/temperature); (2) super-resolution reconstruction of low-resolution fields; and (3) denoising of high-noise measurements. Validation on a steady laminar Bunsen flame demonstrates that CRF-PINN achieves high accuracy across all tasks. The performance of CRF-PINN is also validated for unsteady flame dynamics, where it accurately infers the sinusoidal motion of a Bunsen flame via cross-translation. Furthermore, tests on the Sandia turbulent jet flame confirm the ability of CRF-PINN to effectively reconstruct complete physical fields from sparse experimental data. CRF-PINN bridges the gap between experimental cost and data fidelity, offering a valuable tool for combustion analysis and combustor optimization.