In this study, the additive manufacturing of a functionally graded material (FGM) via directed energy deposition was investigated as an alternative to joining dissimilar metals. The metal powder composition of the FGM was gradually changed from fully low-carbon steel to austenite steel along the building direction. A convolutional neural network model was employed to classify the austenite, martensite, and ferrite phases in the FGM. The volume fraction of the phases was calculated using X-ray diffraction Rietveld refinement and compared with that predicted by the thermodynamic model and that determined from electron-backscattered-diffraction maps. The volume fraction of the bcc phase gradually increased, and the grain size decreased from top to bottom. Nanostructural investigations confirmed the absence of carbide and twin structures due to the relatively low carbon concentration in the upper layers and the presence of a hexagonal ω-Fe phase with twin structures in the interlayers. Furthermore, electron channeling contrast images and kernel average misorientation maps revealed the activation of the deformation twinning and strain-induced transformation of the retained austenite to martensite, which increased the strain-hardening rate. This study can guide the selection of a tailored manufacturing strategy and process parameters to obtain the required material distribution.