DistortionML

Machine Learning approach for quantifying detector distortion fields. This project is a feasibility study for training a surrogate model (possibly NN) to represent the distortion inherent to X-ray pinhole cameras using a nearby, divergent source.

Things to do:

• remove the hexrd dependency
  • makea local version detectorXYToGvec
  • replace the use of the instrument module by extracting the necessary parameters directly from the HDF5 config file.
• make a PyTorch implementation of the pinhole_camera_module
• set up a test training problem

Running

This project currently depends on hexrd; the simplest way to get running is to use conda. It is highly recommended to put hexrd into its own virtual env:

conda create --name hexrd python=3.8 hexrd -c conda-forge -c hexrd

For the bleeding edge version of hexrd, the channel spec is

conda create --name hexrd python=3.8 hexrd -c conda-forge -c hexrd/label/hexrd-prerelease

The script compute_tth_displacement.py executes the distortion field calculation based on the single-detector instrument in resources/. It has a progress bar, and plots the distortion field when it completes. You can run it interactively in your favorite IDE, or IPython:

ipython -i compute_tth_displacement.py

Parameters

The editable parameters are all located in the following block at the top of the script:

# =============================================================================
# %% PARAMETERS
# ============================================================================='
resources_path = './resources'
ref_config = 'reference_instrument.hexrd'

# geometric paramters for source and pinhole (typical TARDIS)
#
# !!! All physical dimensions in mm
#
# !!! This is the minimal set we'd like to do the MCMC over; would like to also
#     include detector translation and at least rotation about its own normal.
rho = 32.                 # source distance
ph_radius = 0.200         # pinhole radius
ph_thickness = 0.100      # pinhole thickness
layer_standoff = 0.150    # offset to sample layer
layer_thickness = 0.01    # layer thickness

# Target voxel size
voxel_size = 0.2

The most sensitive parameter is voxel_size, which essentially will set the size of the problem, since the number of evaluations will increase quickly for increasing voxel size. Making layer_standoff larger will also increase the total number of voxels contributing for a particular voxel_size.