From Synthesis to Kinetics: A Data-Driven Deep Learning Framework for Process-Aware Ferroelectric Dynamics

Physical modeling of emerging ferroelectric HfxZr1−xO2 (HZO) materials is constrained by incomplete theoretical descriptions of domain kinetics and the difficulty of formulating efficient equations for complex evolutionary mechanisms, such as the wake-up effect. To address this challenge, we propose a data-driven Deep Learning framework utilizing an optimized Residual Artificial Neural Network (ResANN). By incorporating critical fabrication parameters as direct inputs, our model learns the intrinsic process-property mappings directly from experimental data. While standard ML-based approaches often leverage idealized simulation data, this work expands the scope by rigorously training on large-scale device measurement data, thus establishing a direct bridge between fabrication parameters and material behavior. By leveraging an optimized deep learning framework and Transfer Learning, our methodology demonstrates exceptional capability in capturing highly non-linear polarization dynamics. We show that this approach yields superior accuracy (Adjusted R2 ≈ 0.998) compared to traditional physical equations, effectively predicting hysteresis evolution and achieving zero-shot generalization to untrained material thicknesses. Consequently, this framework serves as a predictive tool for inverse material design, guiding the optimization of synthesis protocols for next-generation ferroelectrics materials.