Shadowgraph data holds significant potential, as it incorporates depth-of-field effects, enabling the extraction of richer physical information [1, 2]. The integration of physical information into it using Physics-Informed Neural Networks (PINNs) has been shown to successfully reconstruct flow fields in compressible, inviscid flows [3]. In this study, we aim to extract 3D field information from experimental shadowgraph data of turbulent convection using a trained deep learning model. Specifically, we em- ploy a U-Net architecture to predict the temperature field over a 2D slice from shadowgraph. However, ground-truth field data is often unavailable or incomplete from experiments. Therefore, as first step, we use numerical shadowgraphs (simply modeled as Laplacian of temperature field) as input and tempera- ture field as output from direct numerical simulation (DNS). Preliminary results, presented in Figure 1, highlight the potential of this approach. Furthermore, we investigate the feasibility of simultaneously ex- tracting velocity fields alongside temperature fields. Finally, we apply the trained model to experimental shadowgraph images as DNS compliments the experimental conditions.