Calculating Temperature and Species Concentration from Quasi-2D Combustion Video Data using Physics-Informed Neural Networks

The field of complex reacting flows occupies a unique academic niche, characterized by partial theoretical knowledge and imperfect data measurements. Physics-Informed Neural Networks (PINNs) bridge this gap, fusing limited theoretical understanding with incomplete datasets to construct a more comprehensive picture. This is particularly relevant in the study of quasi-2D combustion processes, which are prevalent in high aspect ratio environments such as car engines. These systems are markedly affected by a spectrum of instabilities—including thermal expansion, buoyancy-driven movements, viscous effects, and diffusive-thermal phenomena—that collectively impact the essential parameter of flame speed. The experimental data for such studies are best collected using a Hele-Shaw apparatus, which involves two parallel plates separated by a small gap, allowing for the observation of these instabilities in a controlled environment. Our study focuses on a dilute hydrogen flame observed through this apparatus, with particular attention to the cellular instabilities and unique sawtooth shapes formed under specific conditions. The video analysis component is critical, as it challenges us to discern meaningful data from high-frame-rate videos, where visibility issues arise due to the diminished signal-to-noise ratio at increased frame rates. This problem underscores the need for sophisticated processing techniques to enhance video clarity and extract valuable data. By applying PINNs to this video data, we have constructed detailed profiles of temperature, species concentration, and velocity, offering a new lens through which to view combustion dynamics. This methodology not only allows for the deconstruction of video data into a richer dataset but also offers a lower-cost alternative to traditional experimental and computational approaches. The continuous nature of the function provided by PINNs, as opposed to discrete datasets, presents a significant advantage in analyzing these complex systems.