Stirred tank reactors are essential unit operations in chemical and biotechnological processes, where accurate models of the flow field are crucial for process design and scale-up. In such multi-query scenarios, efficient surrogate models are required. Physics-informed neural networks (PINNs) represent a promising approach to reduce both the computational cost, compared with high-fidelity CFD simulations, and the cost of generating training data from simulations or experiments by embedding the governing equations in the learning process. However, “vanilla” PINNs exhibit several well-known failure modes, including difficulties in capturing high-frequency solution features, sensitivity to sampling, and imbalanced contributions of different loss terms. To address these issues, a variety of approaches have been proposed, such as adaptive loss-scaling strategies, adaptive sampling, and domain decomposition. In this work, we investigate and compare the impact of these techniques on the prediction accuracy and training cost of a PINN-based surrogate model for the 3D flow in a stirred tank reactor, modeled using the RANS equations and a multiple reference frame approach for impeller rotation. We show that transferring successful PINN configurations from 2D to 3D is not straightforward, analyze the trade-off between accuracy gains and additional computational effort, and discuss the minimal data requirements for constructing reliable 3D surrogate models of flow in stirred tank reactors, in comparison with classical supervised neural networks.