Skip to main content
Press Release | Mathematics and Computer Science Division

Using artificial intelligence to detect blobs

New tool incorporates a deep-learning-based approach for accurate blob detection in fusion simulation data

The mention of blobs may bring to mind something we painted in preschool and our parents hung on the refrigerator door. Or the word may call up a vision of an alien creature featured in a horror movie. Why would a scientist want to study blobs? Even more intriguing, how would artificial intelligence be involved — unless, of course, one wanted to avoid getting too near a blob.

Blobs are fascinating — and dangerous,” said Hanqi Guo, an assistant computer scientist in the Mathematics and Computer Science (MCS) Division at Argonne National Laboratory. They are regions of high turbulence that run along the edge wall of tokamak reactors and are believed to be a major cause of the intensely hot plasmas losing heat at that edge.”

Scientists have used simple methods to identify the contours of these blobs in fusion simulation data, but the methods are based on arbitrary choices that involve manual selection.

To address this problem, a team of researchers from Argonne, the University of Notre Dame, and Princeton Plasma Physics Laboratory developed ContourNet, which incorporates a deep-learning-based approach for blob detection.

The researchers first inspected 2D cross-sections of the reactor at a given time step. They then worked closely with domain scientists to label blobs in these cross sections.

We also co-designed an automatic label propagation strategy that extends the labeled data to neighboring time steps. This strategy enabled us to generate 10 times the number of labeled blobs originally obtained,” said Tom Peterka, a computer scientist in Argonne’s MCS Division.

Although the strategy is an approximation, the propagated blobs were verified by the scientists as acceptable for training purposes. All the labels — both those from the scientists and those from propagation — were then used to train ContourNet for 100 epochs. The results showed that ContourNet can achieve high accuracy, correctly identifying relevant potential blob candidates across a large spectrum of different simulation data.

One possible limitation was the identification of false positives in some cases. But false positives are better than false negatives, in that they provide material for the experts to examine.

 This deep learning strategy helps minimize the burden of manual labeling and maximize the blob detection accuracy,” Guo said. Using ContourNet, scientists can better understand the blob dynamics influencing the confinement of the plasma.”

For further information, see the paper by Martin Imre, Jun Han, Julien Dominski, Michael Churchill, Ralph Kube, Choong-Seock Chang, Tom Peterka, Hanqi Guo, and Chaoli Wang,
ContourNet: Salient Local Contour Identification for Blob Detection in Plasma Fusion Simulation Data,” in ISVC19: Proceedings of International Symposium on Visual Computing, pages 289301, Lake Tahoe, NV, 2019.