Being able to produce custom plastic parts on demand has been transformative for industry and hobbyists alike. But 3D printing is not just about plastics anymore — materials like concrete, metal, wood, and also conductive and magnetic materials have opened up new worlds of possibilities. Unfortunately, printing with new materials can be very difficult. Understanding how to get successful prints from a new material is typically a time-consuming and costly process of trial and error. It can take thousands of prints to determine the correct parameters, like print speed and material deposition rate. To make matters worse, ideal parameter values can change with varying environmental conditions and between batches of material.
Problems such as these can keep many from experimenting with using alternative materials in 3D printing, losing out on all of the possibilities they provide. There may be a better path forward coming on the horizon, however. Researchers at MIT have developed a method to simplify the process of choosing optimal 3D printing parameters with the help of machine learning. The system uses computer vision to monitor prints in real-time and can even correct prints on the fly before they have a chance to go astray.
First, the team needed to get the hardware configured, so a 3D printer was outfitted with a pair of cameras aimed at the extruder. Lights are then shined at the material as it is deposited so that the amount of light passing through can be captured by the cameras to serve as a proxy for the material’s thickness. That data is fed into a neural network, in real-time, that is capable of making predictions about the best parameters to use for the print. Those predictions can then be used to tune the printer as it progresses with the print.
Before the network was ready to be used, it of course needed to be trained. The team decided on a reinforcement learning-based approach, in which the model learns through trial and error. When a set of parameters produces an object that is very close to the expected output, the model is rewarded so that it knows it is on the right track. But to achieve a high degree of accuracy with a model like this, it would need to be shown millions of prints.
It would seem that the researchers did in fact have some plans for their lives other than training this model, so rather than create all of these physical 3D prints, they opted to do it in a simulated environment. The real world is not quite so nice and tidy as what might be expected from a computational simulation, however, so they first needed to inject some noise into the simulation with a numerical model that closely approximates the noise seen with a real-world 3D printer. This yielded very realistic data, and allowed it to be collected very quickly.
With the hurdles having been cleared, it was time to check and see if all of this work actually panned out in reality. This new machine learning-powered printer controller was compared with some traditional controllers. The new technique was found to produce more accurate results, and it was noted that it performed especially well when it came to infill printing. Existing methods often deposited too much material, sometimes to the point where the test object bulged up and was ruined, but the new method adjusted parameters and kept the print on the right track.
The controller was even found to be capable of adjusting to new materials in the field, without needing to revalidate the manufacturing process. With the successes that they have seen so far with 3D printing, the team is now exploring how they can apply their innovation to other manufacturing processes. They also believe that there is room for improvement on the 3D printing side of things as well — they are presently exploring how they can use their method when multiple materials are being printed at the same time.