The thermal history during additive friction stir deposition (AFSD) preheating affects subsequent deposition process and material properties, but its prediction depends on an appropriate heat source model. This study presents a heat source reconstruction method using a multilayer perceptron (MLP). A finite element model was established to represent frictional heat generation at the feedstock-substrate and deposited layer-tool shoulder interfaces, together with volumetric heat generation caused by plastic dissipation [1]. The unknown model parameters were calibrated against published experimental temperature data [2,3]. Surface heat flux density and volumetric power density distributions were extracted separately onto regular grids and used to train two MLP models that map spatial coordinates and time to heat source power density. The relative difference between the total power obtained by gridded extraction and direct integration remained below 10 % during most of the preheating stage. On the test set, the relative errors in the total power predicted for both heat sources were below 10 % after 2 s. The reconstructed fields reproduced the temporal expansion of the heat source region and the gradient charcteristics of the reference distributions. The main reconstruction errors occurred near heat source boundaries and during the initial transition between different heat generation mechanisms. These results demonstrate that the MLP model can serve as a surrogate for the analytical heat source definitions and provide a basis for extending the model to other processing conditions.