Finding material defects the smart way

Finding defects in materials is important for improving existing materials and creating new ones. A method called coherent diffractive imaging helps scientists see tiny structures like nanocrystals and their flaws. Recently, researchers have started using deep neural networks (DNNs) to make this process faster. However, DNNs need huge, varied datasets to learn from — and creating these sets is time-consuming and requires a lot of computer power. 

To solve this problem, researchers from the U.S. Department of Energy’s (DOE) Argonne and Oak Ridge national laboratories Laboratory created an automated method that works well with smaller datasets.Their study was published in the journal Neural Computing and Applications. 

A smarter approach 

The researchers combined two key ideas. The first is the use of continual learning (CL), which trains the model step by step as new data is observed, rather than all at once. The challenge is to avoid ​“catastrophic forgetting” — overwriting older data with new information. Here the researchers aim for a balance between minimizing forgetting and efficiently learning on new data. 

The second key idea is smart data generation, which focuses on the most promising regions when selecting the next data samples. 

“By combining these two ideas in what we call smart continual learning, or smartCL, we can classify the various defects more accurately,” said Orçun Yildiz, an assistant computer scientist in Argonne’s Mathematics and Computer Science (MCS) division and lead author of the journal paper.  

With smartCL, the researchers label the different types of defects (like point defects). They then generate the next data samples based on a score that shows how useful the samples will be for model training. This process helps the researchers focus on the most critical defects. 

Higher accuracy, less data, lower costs 

Yildiz and his colleagues tested smartCL on the Swing cluster in Argonne’s Laboratory Computing Resource Center. The results, shown in Figure 1, indicate that smartCL achieved high accuracy more quickly than regular continual learning.