Plastic injection molding has become an important technique in traditional industry in recent years. In the process of injection molding, many manufacturers rely on the experiences of skilled workers, rather than utilizing an efficient method to eliminate processing defects, resulting in difficulties in quality control and increased total cost. To solve the problem of defect removal effectiveness, computer-aided engineering (CAE) is often employed, which can eliminate molding defects, through simulation analysis, before manufacturing. However, some unpredictable problems remain during the actual molding, which require the assistance of field technicians. The outcomes of injection molding, which involve injection pressure, cooling time, and warping deformation, have an intricate connection with control factors, which cannot be classified by regular linear programming. Back Propagation Neural Network (BPNN) has excellent predictive ability in solving non-linear problems. It can accurately predict the results after executing a series of training data. This study combined the orthogonal Taguchi Method and BPNN to construct a computing system for predicting the analysis result of CAE, and analyze the influence of multi-layer structure on prediction accuracy. The results showed that using the Taguchi Method to optimize the parameters of BPNN can increase the accuracy of prediction. Using the optimized network parameters can reduce the prediction error of the maximum injection pressure and maximum cooling time to less than 1%. However, there is still an error of 7.26% for the prediction on warping deformation, which demands further investigation of training data. As such, this study employs the technique of second-training process for further reducing the error of warping deformation to 4.29%.