Post-processing methods are widely used to address the issues caused by surface imperfections and bulk defects in additive manufactured materials. In our previous studies, we analysed the effects of different peening-based treatments of shot peening (SP), severe vibratory peening (SVP) and laser shock peening (LSP) on fatigue performance of V-notched laser powder bed fusion AlSi1Mg samples. Herein, the fracture surfaces of failed samples were further analyzed and obtained experimental data were further elaborated by machine learning (ML)-based approach to identify the correlation between residual stress, hardness and surface roughness (all affected by the applied post-treatments) with the depth of crack initiation site and fatigue life of the post-treated samples. ML-based model was developed via a six layer deep neural network (DNN) as well as using stacked auto-encoder (SAE) for pre-training of the used data set. Taking the advantages of SAE, the accuracies of more than 0.96 were obtained for the predicted results. Correlations were obtained by performing parametric analyses and the importance of each input factor was assessed through sensitivity analyses. The obtained results revealed that by enhancing surface hardening and inducing higher compressive residual stresses as well as more efficient surface roughness reduction, deeper crack initiation site and superior fatigue life can be obtained. In addition, it was found that the depth of sub-surface crack initiation had direct relation with fatigue life improvement in the samples.