Abstract:

Background-oriented schlieren (BOS) tomography has emerged as an effective tool for quantitatively visualizing and reconstructing spatio-temporal volumetric thermal flows. Existing solutions rely on capturing multiview snapshots of a volumetric flow and solving a time-resolved inverse problem to reconstruct the flow. In this work, we propose a single view BOS imaging system that, when constrained by partial differential equations (PDEs) that characterize the airflow temporal dynamics, allows us to accurately reconstruct the time-resolved flow field. Our framework leverages an image rendering schlieren loss metric coupled with a physics-informed neural network (PINN) representation of the target flow fields that minimize the residuals of the coupled Boussinesq approximation of the time-dependent incompressible Navier–Stokes and the heat transfer equations. We further investigate a data-driven closure strategy in which effective thermal transport coefficients are learned directly from BOS data, thereby compensating for model mismatch between prescribed molecular properties and the unresolved turbulent trans- port present in the true flow.