Digital twins must contain at a minimum a physical and virtual asset that describes a given process. Within the semiconductor industry, these digital twins are typically constructed in-house and not shared within the larger community due to privacy concerns. This could create stagnation as techniques are not shared within the broader industry. To enable the ability to share information, rather than proprietary data, well established Federated Learning techniques could be used within the semiconductor industry to create a shared virtual asset. More specifically, this involves creating an Artificial Intelligence (AI) model where the training is conducted locally within an organization rather than a central repository where proprietary data could be at risk. Rather than data being transferred off-site, only the learned hyper-parameters are aggregated within a larger community of locally trained models. Furthermore, raw data is mathematically guaranteed to be anonymized during training, through the use of differential privacy, to ensure that proprietary information cannot be reconstructed through a given model. As a result, Federated Learning has the potential to provide unrealized insights into processes within the semiconductor industry as AI model can ingest much more data than a single organization can provide. This talk will focus on the application of Federated Data, Learning, and Analytics to semiconductor assets through the SMART USA Initiative.
