Advanced Machine Learning-Driven Multi-Objective Optimization for Peak Performance in Ultra-Thin a-IGZO Thin-Film Transistors

The relentless pursuit of miniaturization, in accordance with Moore's Law, has driven memory devices such as DRAM and NAND to unprecedented performance levels. However, this trend has encountered fundamental physical limitations, impeding further scaling of conventional silicon (Si)-based transistors and restricting potential performance gains. To address this challenge, novel materials and innovative device architectures are being actively explored. Within the realm of emerging semiconductor technologies, the Indium Gallium Zinc Oxide (IGZO) channel transistor has garnered significant attention due to its inherently low leakage current characteristics. While IGZO offers promising prospects for low-power memory applications, a key challenge lies in simultaneously achieving high field-effect mobility, a crucial factor for enabling fast operation speeds, without compromising its intrinsic electrical properties.<br/>To address this inherent challenge, a novel machine learning (ML)-based approach is proposed for fabricating ultra-thin IGZO thin-film transistors (TFTs) with exceptional performance using the sputtering process for the transistor channel. This approach employs ML algorithms to optimize multiple sputtering parameters, acting as a powerful tool to overcome the limitations of conventional techniques. Notably, this approach addresses the key challenge of simultaneously achieving multiple competing electrical properties in IGZO TFTs. Consequently, this ML-driven method enables the fabrication of TFTs exhibiting outstanding characteristics, including a high field-effect mobility of 36.5 cm²/V s, a near-zero threshold voltage of -0.08 V, and an impressively thin channel below 7 nm. The fabricated TFTs exhibit performance metrics on par with those achieved using Atomic Layer Deposition (ALD), a well-established technique known for producing high-quality, ultra-thin films. This accomplishment suggests the potential for the ML-optimized sputtering process to address the inherent throughput limitations associated with ALD.<br/>Harnessing the power of machine learning, the proposed method revolutionizes the optimization of the sputtering process, effectively eliminating the need for traditional labor-intensive and time-consuming trial-and-error approaches. This transformative approach holds immense promise for accelerating the development of next-generation memory devices with groundbreaking efficiency and speed.