Researchers from the Massachussets Institute of Technology (MIT), Cambridge, USA, have published the paper “Closed-Loop Control of Direct Ink Writing via Reinforcement Learning” discussing the use of machine learning (ML) to monitor and adjust the Additive Manufacturing process in real-time.
Using simulations, the researchers taught a neural network to adjust build parameters in a way which minimised errors before applying this control to an Additive Manufacturing machine. The research has stated that this ML control enabled more accurate builds than the other Additive Manufacturing controls to which the new one was compared, and was said to have performed especially well at additively manufacturing of the interior of an object.
“This project is really the first demonstration of building a manufacturing system that uses machine learning to learn a complex control policy,” stated Wojciech Matusik, senior author, and professor of electrical engineering and computer science at MIT who leads the Computational Design and Fabrication Group (CDFG) within the Computer Science and Artificial Intelligence Laboratory (CSAIL). “If you have manufacturing machines that are more intelligent, they can adapt to the changing environment in the workplace in real-time, to improve the yields or the accuracy of the system. You can squeeze more out of the machine.”
In order to adopt machine learning, the researchers developed a system using two cameras pointed at the nozzle of the Additive Manufacturing machine. This system shines a light at material as it’s deposited and is able to calculate thickness based on the amount of visible light. The controller then processes these images and subsequently adjusted the feed rate and direction of the nozzle to counteract any errors.
To train the Machine Learning network to understand the Additive Manufacturing process would, however, require an untold number of builds. This is what lead the researchers to use simulation in the early stages.
Using reinforcement learning, the model was asked to select build parameters for certain objects in simulated environments. After being shown the output, the model was ‘rewarded’ if the parameters minimised errors between the plan and the build (’error’ here meaning incorrect material deposition or placement). The more builds the model simulator, the more accuracy was exhibited.
To account for noise during the Additive Manufacturing process, the researchers also created a numerical model which approximated the noise from an Additive Manufacturing machine. This model added noise to the simulation, leading to increasingly real-world results.
“The interesting thing we found was that, by implementing this noise model, we were able to transfer the control policy that was purely trained in simulation onto hardware without training with any physical experimentation,” stated Mike Foshey, one of the paper’s co-authors and a mechanical engineer and project manager in the CDFG. “We didn’t need to do any fine-tuning on the actual equipment afterwards.”
He continued, “We were also able to design control policies that could control for different types of materials on-the-fly. So if you had a manufacturing process out in the field and you wanted to change the material, you wouldn’t have to revalidate the manufacturing process. You could just load the new material and the controller would automatically adjust.”
The work was published at the open-source repository arXiv and is available here.