Automated, physics-guided AI framework for asymmetry-aware ferroelectric compact models.

Despite rapid progress in HfO₂-based ferroelectrics, asymmetry-aware characterization and its reflection in compact models remain insufficiently explored. Also, parameter extraction (PE) is still manual and inconsistent, particularly when asymmetric hysteresis and staged FeCAP (ferroelectric capacitor) to FeFET (ferroelectric field-effect transistor) calibration needs to be captured. To address this, we present an asymmetry-aware ferroelectric compact model and a physics-guided neural PE framework that automate this workflow. For FeCAPs, we generate P-V datasets parameterized by ferroelectric film thickness (tFE) and train a Transformer-encoder PE model to infer target Electrical Parameters (EPs), embedding light physics priors and filtering abnormal loops. For FeFETs, we form ID-VG datasets parameterized by gate length (Lg) and train a one-dimensional convolutional neural network–Transformer (1D-CNN–Transformer) encoder with a hierarchical multilayer perceptron (MLP) head network, regularized by a soft physics prior. In verification process using group-blocked test split, the FeCAP/FeFET PE achieves near-unity correlation and sub-5% error across both device types including interpolation/extrapolation tFE/Lg region. The proposed workflow reduces manual effort and inter-operator variance, enabling rapid and stable held-out interpolation/extrapolation generalization within the anchor-calibrated compact-model domain and thereby accelerating compact-model library generation, process design kit (PDK) enablement, and Design Technology Co-Optimization (DTCO) workflows for FeCAP/FeFET technologies.