Two scientists from the University of Cambridge, Douglas Brion and Sebastian Pattinson, have developed an intelligent 3D printer that can quickly identify and correct the printing mistakes that appear in the process of digital manufacturing. Their research paper was published in Nature Communications on 15th of August 2022.
The research team was able to create a ML (machine learning) algorithm which detects and fixes printing errors that can be used with FDM (Fused Deposition Modeling) 3D printers.
The 3D printers are vulnerable to manufacturing errors such as poor flow rate, interlayer defects, warp deformation, or large top surface defects. Although humans can detect some of these errors, they are not able to provide real-time corrections.
The ML algorithms can better perform these tasks by reducing the number of potential print failures and correcting the errors in real-time.
Model architecture, training and performance
The research team used a Deep Learning Computer Vision model trained on a dataset containing 946,286 labeled images.
The images were previously taken from 192 prints of different 2D and 3D geometries. Each image was labeled with information about some printing parameters such as the printing nozzle’s speed and temperature, along with information about how much they deviated from the ideal values for that image.
The software known as CAXTON (collaborative autonomous extrusion network), can link and manage these learning 3D printers, automatically flagging mistakes and making corrections in real time.
With the aim of developing a generalized procedure which can detect an issue regardless of the piece or material utilized, the researchers focused on this autonomous generation of data to produce larger and more diversified datasets (see image below):
Image legend:
a. Workflow for collecting varied datasets from extrusion 3D printers with the automatic labeling of images with printing parameters.
b. Fleet of eight thermoplastic extrusion 3D printers (Creality CR-20 Pro) equipped with cameras focused on the nozzle tip to monitor material deposition.
c. Renderings of generated tool paths for a single input geometry, with randomly selected slicing parameters.
d. Snapshot of data gathered during an example print showing images with varying parameter combinations.
e. Design of bed remover and dock utilizing existing motion system along with photographs taken during operation.
f. Distributions of normalized parameters in the full dataset collected by CAXTON containing over 1.2 million samples.
During the testing stage the trained model was able to detect and correct the printing errors. Unlike humans, the system operated continuously and instantly made corrections.
Despite these advances there are still unsolved problems with real time 3D printing, such as the mechanical errors caused by skipped steps, belt slipping, or external interference. Further research will be required in order to address these issues.
Current trends show a promising future in 3D printer manufacturing, with a wide range of applications and the potential to revolutionize the way products are made.
Learn more:
- Research paper: Generalisable 3D printing error detection and correction via multi-head neural networks (on Nature Communications)
- Adaptive 3D Printing for In Situ Adjustment of Mechanical Properties (on Wiley Online Library)
- Using artificial intelligence to control digital manufacturing (on MIT News)
- The source code used to generate the results in the research paper is available (GitHub repository)
