China3D printingNet October 14th, researchers at Lehigh University in Pennsylvania have developed a new method based on machine learning that can classify groups of materials based on structural similarity.
In what the team believes is the first of its kind, artificial neural networks are used to identify structural similarities and trends in a huge database containing more than 25,000 material microscopic images.The technology can be used to discover new materialsDevelopmentThe research between them can even correlate factors such as structure and attributes, which may provide a new method of computing material development for 3D printing and other fields.
The lead author of the study, Joshua Agar, described how the model’s ability to detect structural symmetry became the cornerstone of the project’s success. He said: “One of the novelties of our work is that we built a special neural network to understand symmetry and used it as a feature extractor to better understand images.”
The illustration of the neural network shows the symmetry image similarity from a database of more than 25,000 piezoelectric response force microscope images. The picture comes from Lehigh University.
The relationship between structure and performance
In materials research, understanding how the structure of a material affects its performance is a key goal. Nevertheless, due to the complexity of the structure, there are currently no widely used indicators to reliably determine how the structure of a material will affect its performance. With the rise of machine learning technology, artificial neural networks have proven themselves as potential tools for this application, but Agar still believes that there are two main challenges to overcome.
First, most of the data generated by materials research experiments has never been analyzed by machine learning models. This is because the results generated (usually in the form of microscopic imaging) are rarely stored in a structured and usable way. The results are often not shared between laboratories, and of course there is no centralized database that can be easily accessed. This is a problem in general materials research, but it is even more so in the field of additive manufacturing due to the larger niche market.
The second problem is that neural networks are not very effective in learning how to recognize structural symmetry and periodicity—the periodicity of material structures. Because these two characteristics are of vital importance to materials researchers, the use of neural networks has faced enormous challenges until now.
Similarity prediction through machine learning
Lehigh’s new neural network aims to solve the two problems described by Agar. In addition to being able to understand symmetry, the model can also search unstructured image databases to identify trends and projection similarities between images. It is achieved by using a nonlinear dimensionality reduction technique called Unified Manifold Approximation and Projection (UMAP).
Agar explained that this method makes it easier for the team to digest the higher-level structure of the data: “If you train a neural network, the result is a vector, or a set of numbers, which is a compact descriptor of features. These features help Classify things in order to learn some similarities. However, the resulting space is still quite large, because you may have 512 or more different features. So, you want to compress it into a space that humans can understand, such as 2D or 3D.”
The Lehigh team trained the model to include symmetrical perception features and applied it to a set of 25,133 unstructured piezoelectric response force microscope images collected over a five-year period at the University of California, Berkeley. Therefore, they were able to successfully combine similar materials based on the structure, paving the way for a better understanding of the structure-performance relationship.
Ultimately, this work shows how neural networks, combined with better data management, can accelerate research on additive manufacturing and materials development in the broader materials community.
Comparison of UMAP projections using natural images and symmetrical perception features. The picture comes from Lehigh University.
China3D printingNet Comments: The predictive capabilities of machine learning have really begun to be used in many aspects of additive manufacturing. Researchers from Argonne National Laboratory and Texas A&M University have previously developed an innovative method to detect defects in 3D printed parts. Using real-time temperature data and machine learning algorithms, scientists can establish a correlation between thermal history and the formation of underground defects.
Elsewhere, in the commercial sector, the engineering company Renishaw has partnered with 3D printing robotics specialist Additive Automations to develop deep learning-based post-processing techniques for metal 3D printed parts. The partnership involves the use of collaborative robots (cobots) and deep learning algorithms to automatically detect and remove the entire supporting structure.
China3D printingnetCompile the article!
(Editor in charge: admin)
0 Comments for “Researchers use new image-based machine learning techniques to improve material development”