An optimization method based on feature extraction and machine learning is provided. At least one input parameter is received. Multiple first historical mold data are retrieved. A similarity calculation is performed according to the input parameter and the first historical mold data. Multiple candidate mold data are selected according to the similarity calculation. The mold design parameters of the candidate mold data corresponding to each input parameter are replaced by the input parameter, and multiple first representative mold data for performing a first simulation analysis are generated. Multiple key feature parameters are found, and multiple second historical mold data are retrieved according to the multiple key feature parameters. An expected data is found, and the mold design parameters of the expected data are filtered and optimized to find multiple second representative mold data for performing a second simulation analysis. At least one set of mold production parameters is generated.