Radiofrequency-induced heating is a major concern in MRI and these risks widely vary between patients. Therefore, we propose a personalized, deep learning temperature estimation method. After electromagnetic and thermal simulation on two human models in a radiofrequency coil, we trained a neural network on one brain slice to predict internal brain temperatures, using the following features: tissue properties, distance to four surface sensors and the corresponding four surface temperatures. Fast testing performed on both intra- and inter- brain slices revealed similar thermal maps compared to simulated maps. Ongoing work targets a better generalization to different anatomies and in vivo experiments.