Reinvent Manufacturing—From Design to Delivery

In semiconductor fabrication, lithography using UV light is one of the key patterning steps to transfer patterns on photomask to photoresist on wafer. At sub-7nm node, EUV (extreme UV, with 13.5nm wavelength) lithography is used at the layers with the smallest critical dimension. The quality of the photomask is of paramount importance, as any defect on the photomask is repeated on all wafers, severely impacting the yield. The standard way for photomask quality control is through mask inspection, a step in which optical images of the mask are collected using a mask inspection tool with the same (actinic) or different (non-actinic) wavelength. Actinic inspection is more desirable as it can capture phase defects, and is the focus of this study. A novel approach combining a compact machine learning model with a traditional optical simulation accelerated by GPU for reference image generation is presented for high-volume die-to-database EUV mask inspection.

The future smart factory is based on Industry5.0, which is adding concepts such as human-centric (helping workers with technology) and resilient (lead time reduction) to the extension of the 4th industrial revolution. Several key concepts are (1) multi-sourced data transmission, storage, and analysis; (2) individualized human-machine interaction; and (3) virtual space and object simulations. These future smart-factory concepts can be realized by a "metafactory" platform, driven by cloud and digital twin technology. For those who want to start a manufacturing business or upgrade a production line, this metafactory platform service will change the whole factory life cycle. Before building a real factory, a virtual factory is built and used for production optimization, control software virtual commissioning, and remote operation.

Physics-based simulations are routinely used to help optimize designs and understand sensitivities with respect to input design variables. Despite significant increases in simulation throughput made possible in part by GPU acceleration, brute-force search for optimal designs using ensembles of detailed calculations can still be untenable due to prohibitive computational costs. Alternatively, adjoint-based simulation approaches or simple response surfaces are utilized to extract necessary sensitivity information in a more computationally efficient manner, but both are challenged when the design objectives are highly nonlinear or potentially discontinuous. We'll demonstrate using a machine-learned model to extract design sensitivity information from a limited subset of high-fidelity flow simulations.