There is great interest in “digital twins” to improve many aspects of semiconductor manufacturing, from increased device yield and performance, reduced consumption of energy and materials, increased flexibility, and to enable rapid uptake and scaling of new material, equipment, and process innovations. The digital twin has both physical and virtual components, with bilateral communication and control; the hope is to enable a wide range of models (of equipment, processes, wafers) at different fidelities (physical to simplified empirical, and machine-learning enabled), to support a wide range of “smart” functionalities. The road to digital twins goes through and builds upon many well-trodden paths. Here, several lines of research at MTL since the late 1980’s are highlighted, beginning with elements of the MIT Computer Aided Fabrication Environment including process flow languages, to DOE/Opt methods for automated surrogate model construction, and run by run control to track and compensate for equipment state and wear in CMP and other unit processes. The development of “statistical metrology” methods encompassed characterization and modeling of semiconductor variation, with layout pattern dependent models to identify “hot spots” in planarization, dishing, and erosion for a given design, as well as to guide dummy fill generation. An evolution from statistical to ML/AI approaches, particularly Bayesian methods, enabled design for manufacturability (DFM) for rapid MOSFET characterization, and then rapid fabrication process tuning, as well as AI-enabled anomaly detection. These and other paths bring us to an exciting next stage of the journey: by harnessing advances in sensing and data collection, AI methods, and computational power not possible at the beginning, the community is poised to create and deploy digital twins for semiconductor manufacturing.
While building wearables to measure emotional stress, we learned that deep brain activation during seizures could show up as a change in electrical signals measured on the wrist. This unexpected finding led us to develop a wristband, “Embrace” that today is worn to alert to neurological events that might be potentially life-threatening. This talk will tell the story of Empatica’s development of a product that wins design prizes for its appearance, looks like a cool consumer timepiece, and yet is collecting clinical quality data and running analytics based on sophisticated machine learning to advance personalized health.
The design, testing, and processing of metals is becoming increasingly driven by computation and automation—for instance, gaps in physical models are addressed by machine learning, and additive manufacturing is crossing from prototyping to production. These developments foreshadow a digital transformation in the manufacturing of metal components and structures, optimizing performance across scales, from atoms to meters.
Can data series from a broad patient population be relevant and reliable tools in predicting individual outcomes when compared to personal wellness sensor data? Or, simply put from a patient perspective, “Can what happen to them, happen to me?” Retrieving and making use of “like-me” signal data based on similarity presents challenges far beyond digital marketing’s effectiveness in making targeted book and movie recommendations. By investigating and understanding those unique challenges, our research group has developed an approach based upon locality sensitive hashing (LSH). We will provide an update on our progress towards adapting LSH for fast and accurate Signal Like-Me capability.
2016 MIT Digital Health Conference
Principal Investigator Sherry Turkle