A research team at Huazhong University of Science and Technology (HUST), led by Professor Tianyou Zhai, has developed an
Ag/Sb2O3/Au molecular-crystal memristor array with brain-inspired computing capabilities, designed to accelerate electric grid inspections while dramatically reducing energy consumption.
The devices employ Sb2O3 molecular crystals as the resistive switching layer, where metallic filaments form within naturally occurring molecular cages. This unique mechanism enables ultra-low-power switching down to the femtowatt level, providing an energy-efficient hardware substrate for real-time edge intelligence.
The memristor array supports both non-volatile and volatile modes, allowing dynamic reconfiguration for tasks such as image convolution, edge enhancement, and transient event detection. Array-level tests demonstrate 100% device yield, and stable multilevel retention for over
14 days, confirming its robustness for deployment in real-world inspection systems.
Leveraging a fully integrated hardware platform, the team enabled analog-domain image convolution to be performed directly on the memristor array.
- High fidelity (<0.5% relative error) between hardware-computed currents and ideal values.
- Reliable detection of edge width, gray-level amplitude, and geometric ratios, enabling quantitative defect analysis.
- 97% feature extraction accuracy compared to software processing.
- Per-convolution energy consumption as low as 132.84 fJ and 67.75 TOPS/W energy efficiency, far surpassing commercial AI chips.
By shifting computation to the inspection terminal, this memristor array greatly reduces the need for transmitting high-resolution images to cloud servers, minimizing network load, latency, and overall system power consumption.
Beyond electric grid inspection, the technology shows strong potential for autonomous drones, industrial IoT, robotics, and low-power environmental monitoring, providing a new path toward next-generation edge AI hardware capable of operating in data-intensive and energy-constrained environments.