Semiconductor front-end fabs perform hundreds of manufacturing steps to produce power electronic devices using various equipment, such as e.g. epitaxy, lithography, and thermal processing tools. Every step is associated with a set of Unit Process Instructions (UPIs) specifying tool configuration parameters, required materials, and operator actions needed to achieve desired outcomes. For example, an epitaxy UPI determines the growth of specific crystalline layers on a wafer. These UPIs are organized into sequences (process flows) describing the entire manufacturing process for a given gallium nitride (GaN) or silicon carbide (SiC) product. Optimization of UPIs for particular products and tools, and harmonization of UPIs across various process flows, is critical to improving yield, reducing energy/material consumption and waste, achieving sustainable manufacturing practices, and a faster time to market for power electronic applications.
The ATRIA project aims to address these challenges for GaN epitaxy tools with hybrid AI that combine heuristic search and physics models with deep learning and reinforcement learning approaches. The approach will allow the consortium to: (1) research digital twin creation methodologies based on articulating and formalizing domain knowledge and semantics of physics-informed processes; and (2) develop novel planning methods to enhance existing and define new UPIs, as well as to harmonize them across process flows.
ATRIA will develop high-fidelity digital twins of epitaxy tools by leveraging hybrid AI methods, including semantically enriched knowledge models that support physics-informed learning and combine domain knowledge with data-driven models. These digital twins will help to research ways to simulate manufacturing processes with high precision, enabling virtual experimentation and optimization without disrupting real-world production lines.
The focus is on optimizing UPIs for epitaxy tools by combining heuristic search with deep reinforcement learning. This hybrid methodology will enable the efficient exploration of high-dimensional parameter spaces, guiding the search for optimal tool configurations and process parameters to enhance yield, reduce costs, and improve resource efficiency.
ATRIA targets to harmonize process flows by analyzing and standardizing UPIs across different scenarios. This involves the integration of automatic planning methods with deep learning for the process flow design to minimize reconfiguration costs, improve consistency, and enhance the overall efficiency of manufacturing operations.
The ATRIA outcomes will include a proof-of-concept system (up to TRL 4) demonstrating the feasibility of hybrid AI for optimizing GaN epitaxy processes, thus ultimately contributing to the advancement of semiconductor manufacturing in this highly competitive market. Sophisticated digital twins for virtual experimentation will reduce the need for costly physical trials and support sustainable manufacturing. The expert-centered design of optimization and planning algorithms will ensure transparency of obtained solutions by integrating domain expertise with advanced AI techniques. The project strengthens international networking by fostering collaboration between Austria and Canada, advancing AI research and knowledge transfer in a transnational context.
