Numerical and experimental approaches are becoming more complementary approaches to address highly turbulent regimes thanks to improvements in computation resources. Nonetheless, experimen- tal results, such as meshless 3D3C Particle Tracking Velocimetry, are often spatially sparse, somewhat discontinuous, subject to measurement noise and incomplete, while CFD simulations still require in- tensive resources to solve all the time scales. To address these issues, we employ Physics Informed Neural Networks (PINNs) [2], which provide a meshless way to enrich and complement experimental data [1]. In the case of turbulent convection, our framework allows 3D temperature discovery, denois- ing, and generating continuous field representations [3]. Departing from the idealized case of Eulerian training-reconstruction, we propose a spatio-temporal sampling framework to tailor our DNS database to closely resemble experimental measurements. This poses additional challenges to PINNs particularly on addressing spatial gaps, tracking loss and scarcity of experimental-type labels. An important ques- tion arises on the optimal choice of collocation points to help alleviating these constraints. Our analysis focuses on the impact of PDE collocation points density through parallel GPU computations. Addition- ally, we explore the effectiveness of smart adaptive spatial PDE sampling to mitigate information loss with respect to relevant turbulent quantities caused by inherently noisy and sparse data.