Heating risks may prevent a patient from receiving an optimal scan. A novel method of internal temperature prediction is proposed by training a neural network on a temperature map from a simulated brain slice, and testing on two other structurally different brain slices. The features are the tissues properties, distance to surface sensors, surface temperatures. The network provides similar maps compared to simulated maps. Current and ongoing work includes optimizing the network parameters to balance the accuracy of predicted temperature with the ability to generalize for the brain anatomy. Future work involves training on two models and testing on others.