Bayesian Dropout Approximation of Process Outcomes
Project Lead: Rensselaer Polytechnic Institute
Member % Cost Share: 50%
CESMII % Cost Share: 50%
Duration: 18 Months
The lack of robust, automatic predictive modeling is a gap that cuts across multiple industries. Traditional data-driven modeling techniques can fail to capture critical elements of processes and require significant domain expertise to exploit for process optimization and control.
- Provide the CESMII Platform with a modeling engine with sophisticated predictive capabilities that can be invoked by Platform apps to model a variety of manufacturing processes.
- Demonstrate the capabilities of that engine via a sensitivity analysis approach on a set of critical semiconductor manufacturing process data for electrical timing.
This project will create a app, containerized and deployed on Azure, and primarily used by Platform apps as a modeling engine. It will interoperate with the SMIB to manage raw and processed data.
The app implements a stochastic neural network approach to capture complex relationships between data features in the form of output distributions.
The app implements a basic suite of tools to postprocess distributions, to extract useful information and visualizations from model outputs.
Key Tasks & Milestones
• API/Deployment definition (Q1)
• Verification test suite definition and construction (Q1-Q2)
• Neural network code (Q2-Q4)
• Code to package trained models (Q3)
• First tier of model sampling tools (Q4)
• First tier of post-processing/visualization tools (Q4)
• Second tier of sampling and post-processing/visualization tools
• Verify all code against test suite (Q4)
• Train network on IBM data set (Q4)
• Execute analyses of IBM data sets (Q5)
Documentation and technology transfer (Q6)
- Applicable to very wide array of processes, with reusable, flexible models.
- Capable of sophisticated predictive modeling
- With low data requirements compared to traditional deep learning
- That captures inherent variations of noisy processes
- That captures “fat tail” behavior of non-Gaussian processes
- That can use dark data sources to increase confidence in predictions of variety of quantities, including yields, resource usage, reliability, catalyst regeneration times
- Capable of sophisticated process sensitivity analyses
- Identify process gains and input interactions, leading to tighter process controls, both run-to-run and online.
- Decrease waste, resource usage.
Addresses Target 2.1.4 from RFP: Modeling and Analytics, Hybrid modeling and application to thermal processes; Addresses Target 2.1.5 from RFP: Smart Manufacturing Platform