Researchers Achieve 10 Billion Write Cycles in Wurtzite Ferroelectric Memory

Researchers from Xidian University, City University of Hong Kong, and Fudan University have demonstrated over 10 billion write cycles in wurtzite ferroelectric memory. The findings, published in the journal Science on September 11, 2026, detail a 100-fold increase in endurance compared to previous limits for this class of semiconductor devices.

The study focuses on aluminum scandium nitride, or AlScN, a material that switches between two electric states to store data. While AlScN is compatible with existing semiconductor processes and offers fast switching speeds, its reliability has historically been limited by the movement of nitrogen vacancies within the crystal lattice.

Wang Ruiqing, a doctoral researcher at Xidian University and co-author of the study, compared these vacancies to missing seedlings in a cornfield. These atomic gaps cluster under repeated electrical stress, forming conductive paths that increase leakage current and eventually trigger device breakdown.

To address this, the team engineered a superlattice structure consisting of alternating thin layers of AlScN and aluminum nitride. These layers act as physical barriers that restrict the movement of nitrogen vacancies, preventing the defects from clustering into failure-inducing paths.

The researchers also implemented a dynamic recovery protocol to restore polarization after heavy cycling. By utilizing partial switching in ultrathin 45-nanometer capacitors, the team reduced the mechanical strain on the lattice per cycle while maintaining sufficient remnant polarization for reliable memory states.

Tests conducted at room temperature reached 4.7 billion cycles under full switching conditions. At 250 K, the devices exceeded 10 billion cycles while keeping remnant polarization above 100 microcoulombs per square centimeter, setting new performance records for the material system.

The approach relies on controlling the defect topology rather than masking the symptoms of material fatigue. By establishing a framework that couples atomic-scale stability to ferroelectric memory, the researchers provide a technical pathway for improving the longevity of nonvolatile storage components.

Modern artificial intelligence systems consume data at rates that expose the endurance limitations of current memory architectures. The ability to integrate nonvolatile memory with high endurance and low energy consumption is a requirement for future in-memory computing designs that aim to minimize data movement between storage and processing units.

This research provides a mechanism for designers to embed memory cells closer to compute logic without the rapid degradation seen in earlier ferroelectric prototypes. By stabilizing the material at the atomic level, the superlattice design allows for more consistent performance in high-frequency, read-heavy workloads typical of large language model inference.

The findings offer a framework for future development, though long-term data retention and thermal stability across wider temperature ranges require further validation. Integration with standard CMOS manufacturing flows will necessitate additional process adjustments to ensure uniformity across 300-millimeter wafers.

The 10-billion-cycle milestone suggests that what appeared to be a fundamental materials limit is solvable through structural engineering. As semiconductor foundries already utilize similar nitride films, the potential for adopting this ferroelectric functionality without requiring exotic new equipment remains a primary area of interest for industry stakeholders.

Further analysis of the data provided in the Science report indicates that the confinement strategy effectively sidesteps the trade-offs between film thickness and voltage requirements. Thicker films previously required higher voltages to achieve stability, but the superlattice design maintains low-voltage operation while providing the necessary structural integrity for extended lifecycle performance.

The research team emphasized that their methodology directly addresses the root cause of fatigue rather than simply applying external corrections. This shift toward intrinsic material stability is essential for the transition of ferroelectric memory from laboratory prototypes to practical, high-density storage applications in commercial hardware environments.