[TOUR] Cornell Nanofabrication Facility (CNF): Machine Learning-Based Process Optimization in the Cleanroom

Fabrication process development for micro and nanostructures requires a substantial amount of calibration and tuning to meet the production yield and device performance metrics. The tuning process to achieve the key metrics needs frequent repetitions and requires a substantial amount of metrology, inspection data collection, and post-processing. The volume and complexity of process data increase with advancing the technology nodes where the traditional methods are inefficient in utilizing the large data volumes. In an academic research facility, often utilizing the legacy production tools and working with larger dimensions, the process data has a more extensive heterogeneity due to the diverse nature of projects, continuous desire to include novel material, and adopt innovative processes. These processes support developing next-generation NEMS devices, photonic structures, micro and nano actuators for robotics, and flexible electronics and quantum nanodevices. Given the wide range of the projects, processes, and materials, the process data, and metrology data in academic cleanrooms, academic facilities are usually not set up with cohesive data infrastructure for the data collection and computing resources for processing at the scale. The lack of infrastructure further limits the utilization of process and metrology data for reducing the cost and time of development. At Cornell Nanofabrication Facility (CNF), we are developing infrastructure to collect the process data and use it with metrology and device performance data to drive new efficiencies in reducing the device fabrication cycle. We use Machine Learning (ML) and Artificial Intelligence (AI) tools to learn to predict a physical process's complex outcome. These include the state-of-the-art semiconductor fabrication and electronic printing processes. A trained model with accurate predictions can reduce the required resources for new process development. In some instances, the dynamic range of AI can be expanded beyond the training database, using the data augmentation and transfer learning methods, to enable assessment of the process window for novel structures and dimensions. We developed an ML (user-defined features) and an AI model (cGAN-based) for the DUV lithography process to predict the process's outcomes. Using the trained DUV lithography model, one can continuously learn the process parameters and reduce the time to optimize a new process. Similarly, for plasma etching, the flow rate of gases and the active plasma power and duration used can be reduced by recognizing events and changes in the etch tool's conditions. Tour – a guided introduction to CNF equipment and the established AI modules will be remotely provided.