The rising demand for low-power, on-device intelligence in edge computing is driving the development of neuromorphic systems that combine novel algorithms with emerging materials and device geometries. Memristive technologies, particularly those based on solution-processed mixed halide perovskites, offer promising routes toward energy-efficient, scalable alternatives to traditional CMOS architectures. In this presentation, we showcase two complementary efforts that advance the use of perovskite memristors in neuromorphic computation.

First, we introduce a large, image-based dataset comprising thousands of experimental current–voltage (I–V) curves from printable perovskite memristors. [1] A convolutional neural network (CNN) is trained on this dataset to recognize and classify seven distinct memristive switching behaviors. This machine learning (ML)-based approach not only automates the assessment of device quality—achieving up to 91% accuracy in binary switching performance classification—but also establishes a foundation for predictive modeling of optimal operation conditions.

Building on this, we address the challenge of hardware-efficient multiclass classification by proposing a novel Outcome-Driven One-vs-One (ODOvO) algorithm, implemented using optoelectronic perovskite memristors as synaptic elements.[2] The light modulation of synaptic weights, fed in our algorithm from experimental data, is a key enabling parameter that permits classification without modifying further applied electrical biases. By integrating the algorithmic advantages of One-vs-One and One-vs-Rest schemes, our method reduces synaptic resource requirements by at least 10× (only 196 synapses) and achieves competitive accuracy on benchmark datasets like MNIST—all while significantly lowering power and iteration costs through light-based modulation of synaptic weights. Consequently, our approach constitutes a feasible solution for neural networks where key priorities are the minimum energy consumption i.e., small iterations number, fast execution, and the low hardware requirements allowing experimental verification.

Together, these studies illustrate how the synergistic integration of device-level advances and algorithmic innovation can pave the way toward scalable, low-energy neuromorphic platforms. The approaches presented offer a path toward practical, experimentally verifiable AI systems that meet the demands of next-generation edge IoT intelligence.