MIT researchers trained a machine-learning model to monitor and alter the 3D printing process in real-time to correct problems.
Although scientists and engineers continually create new materials with unique qualities that can be utilized for 3D printing, doing so can be challenging and expensive.
3D Printing
An expert operator often has to do manual trials and errors, sometimes making thousands of prints, to find the best settings for a new material that consistently produces the best print quality. These variables are how fast the printer prints and how much material it uses.
MIT researchers are now using artificial intelligence to speed up this procedure. They developed a machine-learning system that uses computer vision to monitor the manufacturing process and correct handling errors in real-time. The controller was installed on an actual 3D printer after the researchers used simulations to teach a neural network how to modify the printing parameters to minimize the mistake. Compared to other 3D printing controllers, their method generated more accurate results.
Work skips the step of printing tens of thousands or hundreds of millions of actual products to train the neural network. It may also make it easier for engineers to incorporate new materials into their designs, allowing them to create products with diverse chemical or electrical capabilities. It may also make it easier for technicians to make quick changes to the printing process if the environment or the material changes unexpectedly.
Simulation
They employed a procedure known as reinforcement learning to train their controller, in which the model learns through trial-and-error with a reward. First, the model’s job was to choose the printing settings that would make a specific object appear in a simulated environment. Then, after seeing the expected result, the model was given a prize if the parameters it chose made the difference between its print and the desired result as small as possible.
An “error” in this scenario means that the model either dispensed too much material, filling in places left open or did not dispense enough, leaving vacant spots we should serve. As the researchers made more simulated prints, the model adjusted its control policy to optimize the reward, becoming more precise.
Real-world process
In contrast, the real world is more chaotic than a simulation. In actuality, slight variations or noise in the printing process frequently cause circumstances to change. As a result, the researchers created a numerical model of 3D printer noise. Next, this model was used to simulate noise, resulting in more realistic results.
The controller produced objects with greater precision during testing than any other control method assessed. It did exceptionally well at infill printing or printing an object’s interior. Other controllers deposited so much material that the printed object bulged, but the controller developed by the researchers regulated the printing course so that the object remained flat. In addition to learning how materials spread after being deposited, their control policy can modify parameters accordingly.
Conclusion
Now that they have proven the strategy works for 3D printing, the researchers plan to develop controls for other manufacturing processes. They are also interested in how we may adjust the method to better deal with instances involving multiple layers of material. Also, their approach assumed that each material’s viscosity (or “syrupiness”) was always the same. A later version might use artificial intelligence to detect and account for thickness in real-time.
Source: indiaai.gov.in
