Thermal simulation in urban digital twins enables effective monitoring of surface urban heat islands and supports climate adaptation planning. This paper evaluates machine learning and physics based approaches for this task through a unified validation framework based on 3D point clouds applied to an urban region in Berlin. The framework enables comparison of RandLA Net for 3D point cloud processing, InfraGAN for 2D texture synthesis, and physics based simulation on triangulated mesh geometries. RandLA Net architecture is adapted for thermal prediction and tested with two feature sets: RGB only and RGB with physics derived material parameters. Deep learning methods demonstrate severe spatial overfitting: training errors are minimal (MAE less than 1 K), but test performance degrades significantly on unseen regions with MAE increasing by factors of 1.9 to 2.5. Unexpectedly, augmenting with material parameters worsens generalization, indicating inadequate feature integration. Physics based simulation maintains consistent predictions (MAE approximately 8 K) with systematic bias addressable through calibration. These results motivate hybrid approaches embedding physical constraints into neural architectures for robust urban thermal modeling.