MACHINE-LEARNING ASSISTED COMBINATORIAL DISCOVERY OF CHALCOGENIDE PHASE-CHANGE MATERIALS

Abstract

Human civilization is entering an era defined by data. Artificial intelligence, autonomous robotics, self-driving vehicles, and ubiquitous sensing networks are generating and processing information at unprecedented scales. Yet, the hardware that underpins modern computing largely based on conventional materials, is struggling to keep pace with this explosive growth in computational demand. Unlocking the next generation of information technologies requires fundamentally new materials that can enable higher speed, lower energy consumption, and novel device functionalities. Accelerating the discovery of suchmaterials has therefore become one of the central challenges in modern materials science and device engineering. This work focuses on leveraging combinatorial high-throughput experimentation and machine-learning (ML) techniques to accelerate the discovery of novel chalcogenide thin films, particularly phase-change materials (PCMs). Chalcogenides—compounds containing Te, Se, or S—exhibit rich structural and electronic properties and have emerged as important materials for applications including two-dimensional van der Waals systems, superconductors, optoelectronics, thermoelectrics, ferromagnets, and phase-change devices. However, the compositional design space of chalcogenide materials is vast, often spanning ternary or quaternary alloy systems, where small variations in composition can significantly alter material properties. Exploring this multidimensional space using conventional approaches is therefore slow and inefficient. To address this challenge, we developed integrated methodologies that combine combinatorial synthesis, high-throughput characteriza- tion, and ML analysis to accelerate materials discovery and identify promising candidates across large compositional spaces. ternary or quaternary alloy systems, where small variations in composition can significantly alter material properties. Exploring this multidimensional space using conventional approaches is therefore slow and inefficient. To address this challenge, we developed integrated methodologies that combine combinatorial synthesis, high-throughput characterization, and ML analysis to accelerate materials discovery and identify promising candidates across large compositional spaces. This thesis begins by incorporating Si into the well-known PCM Sb2Se3, initiating the design of a new PCM family. Chapter 3 demonstrates how combinatorial approaches can be systematically implemented in a pseudo-binary system to explore nanocomposite engineering. The resulting Si–Sb2Se3 system was investigated for its optical properties and microstructures at the thin-film level and further integrated into multiple device platforms to achieve active switching cycles. In Chapter 4, we develop a wafer-scale, on-chip automated characterization platform that enables in operando investigationof material optical and thermal properties on photonic integrated circuits (PICs). The platform is demonstrated using a potential PCM system consisting of mixtures of well-studied Ag–In–Sb–Te (AIST) and indium tin oxide (ITO). Finally, we extend these methodologies toward autonomous experimentation (AE). By integrating multiple instruments within a co- ordinated architecture, the platform replaces conventional sequential characterization with real-time learning of structure–composition–property relationships. Chapter 5 showcases this autonomous framework on the complex Mn–Sb–Te ternary system, where we further identify its potential as a magnetic PCM platform. By developing new data-driven platforms, this work bridges the gaps between thin films and devices, synthesis and characterization, and software and hardware. Most importantly, it accelerates the discovery and optimization of materials for next-generation information processing. Through open-source code and modular system design, the platforms we developed aim to democratize advanced materials discovery and provide a versatile toolkit that can be adapted to a wide range of complex materials systems.