Dataset of Perovskite Memristive Current–Voltage Characteristics for Pattern Recognition

The ever-increasing number of Internet-of-Thing devices requires the development of edge-computing platforms to address the associated demand for big data processing at low power consumption while minimizing cloud communication latency. Neuromorphic computation is a viable solution to avoid an unsustainable energy cost; however, achieving stable memristive switching is a complex process. Mixed halide perovskite resistive memories are a promising technology that usually requires an extensive experimental characterization procedure till a stable operation mode is reached, resulting in an abundance of data that need to be manually processed. In this study, we create a dataset for pattern recognition based on thousands of images of experimental current–voltage (I–V) characteristics of solution-processed, and thus printable, mixed halide perovskite memristors. We have categorized our experimental data into seven distinct categories of memristive behavior depending on the shape of the I–V curves. A machine learning (ML) approach is implemented using a convolutional neural network (CNN) trained using this image-based dataset. After the training phase, the CNN is able to categorize any new experimental I–V across the seven generic types, while a binary categorization process by splitting the experimental data into those exhibiting good or bad switching characteristics is demonstrated with validation accuracies of up to 91%. Overall, it is shown that this ML-based pattern recognition approach can assist in identifying how many of the tested memristive devices exhibit stable, optimum switching dynamics, and upon expanding the model, it could predict which characterization parameters are most influential toward achieving an efficient device operation.