A Machine-Learning-Based Prediction Method for Manufacturing Tolerances in Multilayer Circuit

Manufacturing tolerances (MTs) have a significant impact on the transmission characteristics of millimeter-wave circuits. Therefore, it is urgently necessary to obtain a cost-effective strategy for predicting MTs. This letter proposes a method that combines the vector network analyzer (VNA) with machine learning (ML) to estimate the MTs in multilayer liquid crystal polymer (LCP) circuits. First, an artificial neural network (ANN) surrogate model is trained to map geometric parameters to electromagnetic (EM) responses using a sampled dataset. Subsequently, the S-parameters measured by the VNA are used as the target function of the genetic algorithm (GA) to inversely determine the MTs of fabricated circuit while ensuring its structural integrity. Experimental validation demonstrates that the prediction percentage error of key geometric parameters is below 10%.