A few-shot learning framework using multimodal rheological features for local prepreg thickness prediction in multilayer PCBs

Abstract

As multilayer printed circuit boards (PCBs) move toward high-frequency and high-speed applications, back-drilling stub control and impedance matching have become critical for signal integrity. Their accuracy is strongly governed by the micrometer-level stability of prepreg (PP) dielectric thickness during lamination. However, PCB lamination involves complex thermal-fluid–solid coupling and highly nonlinear resin rheology. Existing mathematical models and simulation methods therefore remain insufficient for accurate local PP thickness prediction, limiting back-drilling depth setting and impedance matching. To address this issue, this study proposes a local PP thickness prediction model for few-shot learning, referred to as PPTPM. First, a 261-dimensional multimodal rheological feature system is constructed from resin-flow mechanisms, including material properties, physics-informed priors, macroscopic rheological trends and microscopic copper-layer morphology, to quantify local resin flow during PP lamination. Second, representative PCB slices are selected by few-shot intelligent sampling, and heterogeneous stacking ensemble learning, multiscale theoretical-baseline residual correction and Tabular Denoising Diffusion Probabilistic Model (TabDDPM)-based data augmentation are integrated to predict local PP thickness from multimodal rheological features. Experimental results show that PPTPM achieves a root mean square error (RMSE) of 6.794 μm using only nine labeled PCB slices, with lower prediction error than existing thickness prediction methods. The model can be used to predict the layer-wise full-board PP thickness distribution for boards from the same part number and batch, providing quantitative support for back-drilling depth setting and impedance control in high-frequency and high-speed PCBs.