New materials with unique properties that can be used for 3D printing are always under development, however figuring out how to print with these components can be complex.
Often, an expert operator must use manual trial-and-error to determine ideal parameters that consistently print a new material effectively. These parameters include printing speed and how much material the printer deposits.
MIT researchers have now used artificial intelligence (AI) in order to streamline this procedure, developing a machine-learning (ML) system that will uses computer vision to watch the production process and then correct errors in real-time.
The experts used simulations to teach a neural network how to adjust printing parameters to minimize error, and then applied that controller to a real THREE DIMENSIONAL printer. Their system printed objects more accurately than all the other 3D printing controllers they compared it to.
The work avoids the prohibitively expensive process of printing thousands or millions of real objects to train the particular neural network. And it could enable engineers to incorporate novel materials more easily into their prints, which could help develop objects along with special electrical or chemical properties. It could also help technicians change the publishing process on-the-fly if materials or environmental conditions change unexpectedly.
“This project is really the first demonstration associated with building the manufacturing system that uses machine learning to learn a complex control policy, ” says senior author Wojciech Matusik, professor of electric engineering plus computer science at MIT who leads the Computational Design and Fabrication Group (CDFG) within the Computer Science plus Artificial Intelligence Laboratory (CSAIL). “If you have manufacturing machines that are more intelligent, they can adapt to the changing environment in the particular workplace in real-time, to improve the yields or the accuracy of the program. You can squeeze a lot more out of the machine. ”
The co-lead authors on the research are Mike Foshey, the mechanical engineer and project manager in the CDFG, and Michal Piovarci, a postdoc at the Institute associated with Science and Technology inside Austria. DURCH co-authors consist of Jie Xu, a graduate student in electrical engineering and pc science, plus Timothy Erps, a former technical associate with the particular CDFG.
Picking guidelines
Determining the ideal parameters of a digital manufacturing process can be one of the most expensive parts of the process because so much trial-and-error is required. And once a technician finds the combination that works well, those parameters are only ideal for one specific situation. She has little data on how the material will behave in other environments, on different hardware, or if a brand new batch exhibits different properties.
Using a ML system is fraught with challenges, too. First, the scientists needed in order to measure what was happening on the printer within real-time.
To do this particular, they developed a machine-vision system using two cameras aimed at the particular nozzle associated with the THREE DIMENSIONAL printer. The system shines light at materials as it is deposited and, based on how much light passes via, calculates the material’s thickness.
“You can think of the vision system as a set of eyes watching the particular process in real-time, ” Foshey says.
The control would then process images it receives from the vision system and, based on any error it sees, modify the feed rate and the direction of the printer.
But training the neural network-based controller to understand this manufacturing process is data-intensive and would require making millions of prints. So , the analysts built a simulator instead.
Successful simulation
To train their controller, they utilized a process known as reinforcement learning within which the particular model learns through trial-and-error with a reward. The model was tasked with selecting printing variables that would create a certain object in a simulated atmosphere. After being shown the expected output, the design was rewarded when the particular parameters this chose minimized the error between its print plus the expected outcome.
In this case, an “error” means the model either dispensed too much material, placing it in areas that should have been left open, or did not dispense enough, leaving open spots that should be filled in. As the model performed more simulated prints, it updated the control policy to maximize the reward, becoming more and more accurate.
However , the particular real world is messier than a simulation. In practice, conditions typically change due to slight variations or even noise inside the printing process. So the researchers created a numerical model that approximates noise from the 3D printer. They used this model to add noise to the simulation, which led to more realistic results.
“The interesting thing we found was that, by implementing this noise model, we were able to transfer the control plan that was purely trained in simulation onto equipment without training with any kind of physical experimentation, ” Foshey says. “We didn’t need to do any fine-tuning on the particular actual equipment afterwards. ”
When they tested the control, it imprinted objects more accurately compared to any other control method they evaluated. It performed especially well at infill publishing, which is printing the interior associated with an object. Some some other controllers transferred a lot material that the published object bulged up, but the researchers’ controller adjusted the publishing path so the object stayed level.
Their own control policy can even learn how materials spread after being deposited and adapt parameters accordingly.
“We were also able to design control policies that can control with regard to different types of materials moving. So in case you had a manufacturing procedure out in the field and you wanted to modify the materials, you wouldn’t have in order to revalidate the manufacturing process. You could just load the particular new material and the controller would automatically alter, ” Foshey says.
Now that they have shown the effectiveness of this technique for 3D printing, the particular researchers want to create controllers regarding other producing processes. They’d also like to see exactly how the approach can become modified intended for scenarios where there are usually multiple layers of materials, or multiple materials being printed at once. In addition, their strategy assumed each material has a fixed viscosity (“syrupiness”), but a future iteration could use AI to recognize and adjust for viscosity in current.
Additional co-authors on this work include Vahid Babaei, who leads the Artificial Intelligence Aided Design and Manufacturing Team at the Max Planck Institute; Piotr Didyk, associate professor in the University of Lugano in Switzerland; Szymon Rusinkiewicz, the David M. Siegel ’83 Professor of personal computer science from Princeton College; and Bernd Bickel, teacher at the particular Institute of Science plus Technology within Austria.
The particular work has been supported, in part, by the FWF Lise-Meitner program, an European Research Council starting grant, and the U. S. National Science Foundation.
