Machine learning–guided optimization of coercive field in Al1−xScxN thin films for nonvolatile memory

Abstract

This study employs a data-driven machine learning approach to investigate specific ferroelectric properties of Al_1−xSc_xN thin films, targeting their application in next-generation nonvolatile memory (NVM) devices. This approach analyzes a vast design space, encompassing over a million data points, to predict a wide range of coercive field values that are crucial for optimizing Al_1−xSc_xN-based NVM devices. We evaluated seven machine learning models to predict the coercive field across a range of conditions, identifying the random forest algorithm as the most accurate, with a test R² value of 0.88. The model utilized five key features: film thickness, measurement frequency, operating temperature, scandium concentration, and growth temperature to predict the design space. Our analysis spans 13 distinct scandium concentrations and 13 growth temperatures, encompassing thicknesses from 9–1000 nm, frequencies from 1 to 100 kHz, and operating temperatures from 273 to 700 K. The predictions revealed dominant coercive field values between 3.0 and 4.5 MV/cm, offering valuable insights for the precise engineering of Al_1−xSc_xN-based NVM devices. This work underscores the potential of machine learning in guiding the development of advanced ferroelectric materials with tailored properties for enhanced device performance.