The thermal cycles in arc welding are crucial as they determine the metallurgy, residual stresses, and distortions of welded parts. Thermal simulations are often used as a preliminary stage, with two main methods: either a multiphysics approach or an equivalent heat source approach. The latter is used to reduce computation times. This study aims at predicting the thermal field using a metamodel approach and experimental data. A non-intrusive, contactless sensor is used for monitoring the weld pool contour. This contour is defined as the reference experimental data. A compact camera integrated into the welding setup acquires weld pool images for contour detection during GTAW operation. A numerical design of experiments is conducted by varying heat source parameters in purely conductive thermal finite element analysis. The resulting dataset is used for machine learning training. An optimization approach employs a polynomial regression model to estimate the heat source from the weld pool contour, while algorithms like K-Nearest Neighbors (K-NN) predict the thermal field from estimated heat source. By employing this datadriven approach, we expect a significant reduction in computational time to obtain the thermal field from the melt pool contour.