Numerical simulations of polymer melt flow behavior in cavities help predict and optimize injection molding process parameters. However, simulation and actual results may differ because of simplified mathematical models, inaccurate processing conditions, material property settings, and machine aging, among other factors. Therefore, simulated optimal process parameters cannot be directly applied in practice. This study applied machine learning to generate a virtual–actual correction model to improve the accuracy of simulation results, especially the cavity pressure profile, a key indicator of injection‐molding quality for identifying ideal process parameter settings such as filling‐to‐packing switchover time and holding pressure. This method does not require big data for model training to enhance its practicality. Therefore, the correction model is only suitable for specific settings. A set of injection molding machines, molds, and processed materials were used for experimental verification. An autoencoder model was used to extract the features of simulation and actual cavity pressure curves. Then, a multilayer perceptron model was used to determine a relationship between simulation and actual features. The autoencoder was used to decode simulated features into cavity pressure curves. The proposed method was verified with dumbbell‐shaped specimens; the correlation between simulated and actual cavity pressures was greatly improved from 81% to 98%.