AI-driven Inverse Design of Complex Oxide Thin Films for Semiconductor Devices

Abstract

Bridging generative foundation models with non-equilibrium thin-film synthesis remains a central challenge, limiting the practical impact of AI-driven materials discovery on semiconductor dielectrics. Here, we introduce IDEAL (Inverse Design for Experimental Atomic Layers), an inverse-design platform that links generative diffusion models, machine learning interatomic potentials, and graph neural network property predictors with atomic layer deposition (ALD). We demonstrate IDEAL using the Hf-Zr-O system as a stringent benchmark for semiconductor-relevant complex oxides. The platform statistically enumerates thermodynamically plausible structures and constructs a composition-structure-property map. Crucially, it identifies a narrow composition window where low-energy tetragonal and orthorhombic phases cluster, revealing trade-offs between band gap and dielectric response. Experimental validation using atomic layer modulation (ALM) corroborates these predictions, demonstrating predictive guidance under realistic, non-equilibrium thin-film growth. By experimentally closing the loop, IDEAL provides a transferable and generalizable route to the precision synthesis of next-generation semiconductor dielectrics.