Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Bayesian Optimization-Based Run-to-Run Learning Control for 3D Printing
In this project, we are interested in developing and testing run-to-run learning methods to optimize the process performance of a 3D printing process while maintaining desirable mechanical properties of the printed part. We utilize Bayesian Optimization (BO), a data driven optimization method that is suitable for optimizing black-box processes through minimizing a desired cost function that characterizes process performance. A state-of-the-art BO method will be utilized to optimize the cost function while satisfying process constraints for the black-box model, which is the part printed by the 3D printer in our context.
The complete project description is available in the attached document.
3D printing processes need parameter optimization for desirable printed part performance. Process optimization for 3D printing is an active research area with little methodical insight other than experimental design methods. While there are promising results in the literature, experimental design methods are not scalable for new designs, materials, and processes. Our goal in this project is to automate the process configuration and optimization by utilizing a mathematical framework provided by the BO method.
The BO method utilizes an acquisition function to iteratively sample the model at inputs that provide evidence of improving the cost function or improving the understanding of the model's behavior.
This problem, better known as the exploration-exploitation problem (E-E problem), is a core concept in learning-based control and machine learning, where an agent takes optimal decisions using the information available to either exploit the knowledge gathered thus far, or learn more at the expense of not being optimal. The in-house BO method that will be used in this project has a novel acquisition function that approaches the E-E problem while satisfying output constraints.
This project will investigate efficient ways to implement the BO architecture for run-to-run learning control, using the data acquired during the printing of the parts to evaluate an optimal set of parameters to be tested on the next part. The method will be first demonstrated in simulation and then on an experimental setup available at our industrial partners.
The student is expected to develop a solution in a simulation environment, e.g. Matlab, and then perform experiments on the experimental setup at the ETH Honggerberg campus. Additionally, our industrial partner has mechanical testing capabilities for rapid verification of printed part performance. Therefore the proposed solutions can be benchmarked against existing solutions.
3D printing processes need parameter optimization for desirable printed part performance. Process optimization for 3D printing is an active research area with little methodical insight other than experimental design methods. While there are promising results in the literature, experimental design methods are not scalable for new designs, materials, and processes. Our goal in this project is to automate the process configuration and optimization by utilizing a mathematical framework provided by the BO method. The BO method utilizes an acquisition function to iteratively sample the model at inputs that provide evidence of improving the cost function or improving the understanding of the model's behavior. This problem, better known as the exploration-exploitation problem (E-E problem), is a core concept in learning-based control and machine learning, where an agent takes optimal decisions using the information available to either exploit the knowledge gathered thus far, or learn more at the expense of not being optimal. The in-house BO method that will be used in this project has a novel acquisition function that approaches the E-E problem while satisfying output constraints. This project will investigate efficient ways to implement the BO architecture for run-to-run learning control, using the data acquired during the printing of the parts to evaluate an optimal set of parameters to be tested on the next part. The method will be first demonstrated in simulation and then on an experimental setup available at our industrial partners.
The student is expected to develop a solution in a simulation environment, e.g. Matlab, and then perform experiments on the experimental setup at the ETH Honggerberg campus. Additionally, our industrial partner has mechanical testing capabilities for rapid verification of printed part performance. Therefore the proposed solutions can be benchmarked against existing solutions.
The goals of the project are as follows:
- Understand the state-of-the-art methods for parameter tuning in 3D printing, and understand the basics of BO as well as the novel method that will be used in the project;
- Develop the learning problem and perform run-to-run testing simulations to demonstrate the results;
- Test the developed method on the experimental setup to gather data and demonstrate the effectiveness of the proposed approach;
- Design and perform tests for tuning controller gains (this could involve design of experiments or Bayesian optimization).
The goals of the project are as follows: - Understand the state-of-the-art methods for parameter tuning in 3D printing, and understand the basics of BO as well as the novel method that will be used in the project; - Develop the learning problem and perform run-to-run testing simulations to demonstrate the results; - Test the developed method on the experimental setup to gather data and demonstrate the effectiveness of the proposed approach; - Design and perform tests for tuning controller gains (this could involve design of experiments or Bayesian optimization).
Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to guidetti@inspire.ethz.ch, ebalta@ethz.ch, rupenyan@inspire.ethz.ch.
Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to guidetti@inspire.ethz.ch, ebalta@ethz.ch, rupenyan@inspire.ethz.ch.