TSV wafer contains extremely large number of irregularly distributed TSVs, the wafer warpage caused by TSV manufacturing process has adverse effects on subsequent processes and thermal-mechanical reliability. As the size and pitch decrease, the interaction among TSVs and substrate becomes significant, ignoring the TSV layout will introduce significant errors in wafer warpage simulation. The current wafer simulation methods mostly ignoring the TSV layout and only consider the thermal load while ignoring other effects induced by manufacturing processes. Therefore, this paper proposes a machine learning (ML)-based TSV wafer modeling and manufacturing process induced strain prediction method. In terms of modeling, the method first divides the TSV interposer into blocks, each block was represented as a tensor. Then an artificial neural network (ANN)-based material equivalent surrogate model was established to quickly calculate the anisotropic equivalent material properties of each block and the wafer FEA model was rapidly assembled by blocks. In the prediction of TSV strain, an initial TSV strain was first assumed and anisotropic equivalent strains of all blocks are calculated using the ANN-based surrogate model, and they were utilized as loads to obtain the wafer warpage. Subsequently, the input TSV strain was iteratively adjusted until the simulated warpage aligns with the measured warpage, thereby the process induced TSV strain was obtained. As an application case, a TSV wafer embedded with 44 interposers was constructed and simulated, the result validated the efficiency of the proposed method.