Synstor circuits for real-time adaptive intelligent systems

1. Introduction

The prevailing method for building artificial intelligence (AI) systems centers on digital computers based on the Turing model (Turing, 1936). Digital computers (Merolla et al., 2014; Silver et al., 2016) and existing neuromorphic circuits (Hu et al., 2017) can accurately implement pre-determined inference algorithms derived from prior learning. A key drawback of these computational approaches is their inherent inflexibility: they cannot alter their explicitly defined algorithms in real-time to adapt to evolving environments. The computationally demanding learning process usually takes place offline on large, energy-intensive computers, and the resulting AI algorithms are then deployed on power-limited edge computing circuits. While achieving superhuman performance within their training domians, AI systems such as autonomous vehicles (Yang et al., 2024), drones (O’Connell et al., 2022), and large language models (Kalla and Smith, 2023) lack the adaptive capacity of human intelligence and falter in unpredictable scenarios outside their learned boundaries. Attempting to extend their training domians using “big data” from dynamic environments proves expensive, inefficient, and ultimately restrictive (Bailey, 2022), as the resulting static AI algorithms cannot fully cope with the boundless complexity of the real world.

2. Synstor circuits for real-time concurrent inference and learning

We developed a synaptic resistor (synstor) designed to mimic a biological synapse (Chen 2022; Danesh et al., 2019; Gao et al., 2023). This three-terminal device was fabricated by vertically integrating a heterojunction with a Si channel, SiO₂ dielectric, Hf_0.5Zr_0.5O₂ ferroelectric memory layer, and a WO_2.8 reference electrode. Lateral Schottky contacts were formed between the Si channel and TiSi_0.9/Ti input/output electrodes (Lee et al., 2025). Figure 1(a) illustrates a synstor circuit inspired by neural networks. Input voltage pulses ( x ) drive output currents ( I ) through the synstors, performing inference ( I = W ( t ) x ), where W is the conductance matrix of the synstor array. These output currents ( I ) trigger actuation voltages ( y ) on actuators via neuron and interface circuits, modifying the system state ( s ). The deviations of s from targeted states s ^ , s − s ^ , are sensed and converted to input pulses ( x ) by interface circuits (Figure 1(b)). Similar to biological networks, the conductance matrix ( W ) can be modified during concurrent learning based on correlative learning rules d W dt = α z ⊗ x , where α is the learning rate, z denotes voltage pulses triggered at the output electrodes, and z ⊗ x represents the outer product between z and x . The simultaneous inference and learning process in the synstor circuit reduces the objective function F = 1 2 ( s − s ^ ) 2 by adjusting W toward W ^ = arg min W F (Chen, 2022).

OPEN IN VIEWER

Figure 1.

(a) Schematic of a crossbar synstor circuit with synstors connecting M input and N output electrodes. (b) The synstor circuit is integrated with a system.

3. Adaptive AI systems based on synstor circuits

We conducted experiments on drone navigation in complex, time-varying simulated aerodynamic environments (Gao et al., 2023; Shenoy et al., 2022) and morphing wing control in a turbulent wind tunnel (Nathan et al., 2022; Shaffer et al., 2021) to compare a synstor circuit, human operators, and a computer-based artificial neural network (ANN). During the drone navigation with strong winds and obstacles, the ANN repeatedly failed and collided, whereas the synstor circuit, and most human operators successfully reached the target by avoiding collisions. In the morphing wing control to minimize drag-to-lift force ratio, the synstor circuit effectively optimized the wing shape and recovered the wing from the stall, whereas the ANN failed in the complex aerodynamic environment. The synstor circuit and human operators learned four orders of magnitude faster than the ANN in the experiments. The ability of the synstor circuit to concurrently perform real-time inference and learning by dynamically adjusting its conductance matrix ( W ) enabled continuous optimization of the objective function ( F ). In contrast, the sequential learning and inference in the ANN prevented real-time adaptation of its weight matrix to the complex dynamic changing environments. The synstors exhibited significantly lower conductance ( < 60 nS ) compared to transistors (∼ 0 . 1 mS ). The power consumption of synstor circuit (∼200 nW) for concurrent learning and inference was seven orders of magnitude lower than the aggregate power consumption (∼10 W) of the computer executing the learning and inference functions sequentially in the ANN.

Scaling down synstors promises further improvements in energy and area efficiency, potentially leading to an ultra-low power, highly adaptable brain-inspired AI computing platform.

4. Conclusion

AI systems employing scalable synstor circuits, exemplified by applications in drones and morphing wings, demonstrate significantly faster learning, enhanced performance, greater energy efficiency, and improved adaptability compared to computer-based AI. These circuits offer a promising pathway for AI systems to excel in dynamic and unpredictable real-world environments.