Using convolutional neural networks to classify static x-ray imager diagnostic data at the National Ignition Facility

Hohlraums convert the laser energy at the National Ignition Facility (NIF) into X-ray energy to compress and implode a fusion capsule, creating fusion. The Static X-ray Imager (SXI) diagnostic collects time-integrated images of hohlraum wall X-ray illumination patterns viewed through the laser entrance hole (LEH). NIF image processing algorithms calculate the size and location of the LEH opening from the SXI images. Images obtained come from different experimental categories and camera setups and occasionally do not contain applicable or usable information. Unexpected experimental noise in the data can also occur where affected images should be removed and not run through the processing algorithms. Current approaches to try and identify these types of images are done manually and on a case-by-case basis, which can be prohibitively time-consuming. In addition, the diagnostic image data can be sparse (missing segments or pieces) and may lead to false analysis results. There exists, however, an abundant variety of image examples in the NIF database. Convolutional Neural Networks (CNNs) have been shown to work well with this type of data and under these conditions. The objective of this work was to apply transfer learning and fine tune a pre-trained CNN using a relatively small-scale dataset (~1500 images) and determine which instances contained useful image data. Experimental results are presented that show that CNNs can readily identify useful image data while filtering out undesirable images. The CNN filter is currently being used in production at the NIF.