Found in everything from the planes we fly in, to the ships moving goods across the ocean, 3D-printed parts are now commonplace. Although it is now possible to print increasingly complex shapes in a wide variety of materials, certain structures remain out of reach as we design features beyond current process and equipment limits.
For example, while advances in an additive manufacturing process called powder bed fusion have enabled lightweight lattice structures to be printed in metal, such lattice structures often suffer from localized shape distortions, with the degree of distortion increasing with the angle of the overhang desired.
“For sub-millimeter lattices with large overhang angles, limitations arise related to laser spot size and the layer-by-layer printing thickness, which lead to systematic distortion of cross-sectional shapes and high surface roughness,” explained Stefanie Feih, a Senior Scientist at A*STAR’s Singapore Institute of Manufacturing Technology (SIMTech).
As a result, what you design and what is eventually printed can differ significantly, leading to reduced strength and stiffness. Geometric compensation is one method commonly used to minimize such discrepancies. However, current strategies are constrained by their use of pre-defined cross-section approximations—like circles, ellipses or polygons—of the original design shape.
Instead, Feih and collaborators at the National University of Singapore used an artificial neural network (ANN) model to generate lattice designs with free-form cross-sections, enhancing the accuracy of printing by using these cross-sections for improved compensation.
First, the team used 3D printing to create dome lattice structures of various diameters and overhanging angles. Next, they scanned the structures using high-resolution X-ray computed tomography to evaluate systematic deviations from the original design geometry. “This gave us 3D point cloud measurement data that was suitable for training the ANN model,” Feih said. “We used the data without filtering to account for surface roughness, which is an important feature of the model.”
When compared to an established geometric compensation method, the ANN compensation method produced a closer representation of the roundness of the printed cross-section, even for large overhanging angles of up to 60 degrees. The researchers attributed the improved structural accuracy to the ANN’s ability to correct for highly localized imperfections caused by fused powder particles from the supporting powder bed.
Feih said that the study is a “great achievement” for first author Ruochen Hong, a PhD student supervised by Wen Feng Lu at the National University of Singapore, who is working with the SIMTech team on improving the quality and reliability of 3D-printed lattices. Such a compensated design approach could replace costly process optimization studies, she added.
To further improve the technique, Feih noted that future work could focus on automating the process for generating training data for the neural network, given that the current data are material- and equipment-dependent.
The A*STAR-affiliated researchers contributing to this research are from the Singapore Institute of Manufacturing Technology (SIMTech).
