Additive friction stir deposition (AFSD) is a solid-state manufacturing technique that utilizes frictional heat to soften and plastically deform metal materials for layer-by-layer deposition. Like in other metal additive process, the quality of printed products in AFSD is governed by thermal characteristics, which are challenging in describe and predict. This study introduces a novel thermal source model that incorporates the effects of asymmetry between the advancing and retreating sides at the tool-feed rod interfaces, with the microstructural grain distribution of the product serving as a quantifiable reference during the calibration process [1,2].The thermal model is characterized by one-way thermomechanical coupling, incorporating realistic tool force-displacement boundaries and corresponding element activation criteria for accurate reproduction of the additive manufacturing process [3]. A Deep neural network (DNN) was employed to learn the complex correlation between the input heat source parameters and the resultant thermal cycles and grain morphology, which enabled a rapid and precise global search for optimal heat source parameters. Experimental validation for single-layer and mutli-layer bulids indicates a strong correlation in temperature and residual stress. Parameter analysis reveals that in the constructed multilayer 6005A, the temperature on the advancing side is significantly higher than that on the retreating side, and the resulting temperature gradient differences also induce an asymmetric distribution of grain sizes. The proposed cross-scale thermal source model calibration method, grounded in the temperature evolution and the grain distribution in the product, provides a powerful tool for explaining the microstructure and mechanical properties of materials under complex thermal cycles.