Dong, Hongyang; (2023) Machine Learning-Based High-Throughput Processing of Chemical Imaging Data. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PHD_Thesis_HD_Final.pdf - Other Download (32MB) | Preview

Abstract

Experimental powder diffraction patterns have improved dramatically in quality and volume due to advancements in X-ray sources, optics, and detectors over the past decades. During this PhD, we have established a machine learning-based chemical image analysis pipeline to accelerate and enhance the acquisition of chemical imaging datasets. Phase Quantification Neural Network (PQ-Net) is a regression, deep convolutional neural network that provides quantitative analysis of powder X-ray diffraction patterns and has been applied to recently acquired data to accelerate data analysis. The network is trained with simulated datasets and tested with experimental X-ray diffraction computed tomography (XRD-CT) datasets and shows great potential as a tool for real-time analysis of diffraction data during in situ experiments due to its ability to yield results at least ten times faster (i.e. minutes rather than hours). We also used a self-supervised learning approach, the SingleDigit2Image (SD2I) network, to overcome the problem related to insufficient projections. The network is tested with simulated data and experimental synchrotron X-ray micro-tomography and XRD-CT data. Statistical analysis revealed that the results are more accurate than the images reconstructed by filtered back projection (FBP) and iterative algorithms. Based on the SD2I, we developed the ParallaxNet to solve the parallax artefact in XRD-CT images. The parallax artefact arises from the incorrectly recorded diffraction angle caused by the different starting positions of scattered beams, which is the main barrier to applying the XRD-CT to samples on the scale of centimetres. The ParallaxNet is tested with one simulated and two real experimental datasets and correctly refined the diffraction reflections in all pixels. Furthermore, a new 3D image decomposition method Self2Comp is developed to decompose the XRD-CT image with the standard diffraction patterns of chemical components. This method requires no prior knowledge about the sample and can extract more components than conventional image-decomposing methods like non-negative matrix factorization (NMF).

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item