ArCTIc

AlgoRithm for Charge Transfer Inefficiency (CTI) correction

Add or remove image trails due to charge transfer inefficiency in CCD detectors by modelling the trapping, releasing, and moving of charge along pixels.

https://github.com/jkeger/arctic

Jacob Kegerreis: [email protected]
Richard Massey: [email protected] James Nightingale

This file contains general documentation and some examples. See also the docstrings and comments throughout the code for further details, and the unit tests for more examples and tests.

Contents

• Installation
  • Requirements
  • Instructions
• Usage
  • Python example
  • C++
• Unit Tests
• Files
• Documentation
  • Add/remove CTI
  • CCD
  • ROE
  • Trap species
  • Trap managers
  • Python wrapper
• Version history

Installation

Requirements

You have to make sure that the following libraries are installed on your system: llvm, omp, gsl.

• On Linux, you can install them using your distro's package manager e.g. for Ubuntu:

apt install llvm14 gsl libomp5

• On macOS, you can install then using e.g. homebrew:

brew install llvm libomp gsl

Instructions

There are two ways to install arCTIc and its python wrapper:

source [recommended]

You can also download/clone the source code manually and compile it using the provided makefile. For doing so, you have to perform the following steps:

1. Clone or download & unpack source code i.e.

git clone https://github.com/jkeger/arctic.git

2. Install arCTIc C++ core

   • Run make core to compile the C++ code into an arctic executable and libarctic.so dynamic library.
   • You should now get output from ./arctic --demo.

3. arCTIc python wrapper

   • Run sudo make wrapper (sudo only required on MacOS) to create arcticpy/wrapper.cypython*.so
   • Add /***current*directory***/arctic/python to your system variable $PYTHONPATH and /***current*directory***/arctic to another system variable $DYLD_LIBRARY_PATH
   • You should now get output (in python) from import numpy, arcticpy ; test=arcticpy.add_cti(numpy.zeros((5,5)))

   MacOS: requires sudo make wrapper, or equivalently cd arcticpy; python3 setup.py build_ext --inplace.

pypi/pip

An older version of arctic can be obtained via the pip module of your python installation

<!--python3 -m pip install arcticpy # Finds the wrong (cython?) version!!! -->
pip install -i https://test.pypi.org/simple/ arcticpy

(or possibly (this is what Jascha suggests))

<!--python3 -m pip install arcticpy # Finds the wrong (cython?) version!!! -->
python3 -m pip install --index-url https://test.pypi.org/simple --extra-index-url https://pypi.org/simple/ arcticpy

The version is not as frequently updated, but this process automatically downloads the source files and builds/installs the executable, library and module. If you do not have superuser privileges, you must add the --user argument to install it into your local (home) directory instead. Furthermore, on some macOS system, you may have to explicitly set the architecture by adding e.g. ARCHFLAGS="-arch x86_64" in front of the command.

Usage

Python

ArCTIc will typically be used via the arcticpy python wrapper module, which uses Cython to interface with the precompiled C++ dynamic library.

Several libraries have been written to wrap around or simplify the use of ArCTIc on fits images, such as Pyxel and pyAutoCTI. For example, to correct CTI in a Hubble Space Telescope ACS image, the following reads in CCD data, flips quadrants so their readout register is at position (0,0), then calls ArCTIc:

import arcticpy as arctic
import autocti

data_path = "data_path/image_name"

# Load each quadrant of the image  (see https://pyautocti.readthedocs.io)
image_A, image_B, image_C, image_D = [
    autocti.acs.ImageACS.from_fits(
        file_path=data_path + ".fits",
        quadrant_letter=quadrant,
        bias_subtract_via_bias_file=True,
        bias_subtract_via_prescan=True,
    ).native
    for quadrant in ["A", "B", "C", "D"]
]

# Automatic CTI model  (see CTI_model_for_HST_ACS() in arcticpy/src/cti.py)
date = 2400000.5 + image_A.header.modified_julian_date
roe, ccd, traps = cti.CTI_model_for_HST_ACS(date)

# Or manual CTI model  (see class docstrings in src/<traps,roe,ccd>.cpp)
traps = [
    arctic.TrapInstantCapture(density=0.6, release_timescale=0.74),
    arctic.TrapInstantCapture(density=1.6, release_timescale=7.70),
    arctic.TrapInstantCapture(density=1.4, release_timescale=37.0),
]
roe = arctic.ROE()
ccd = arctic.CCD(full_well_depth=84700, well_fill_power=0.478)

# Remove CTI  (see remove_cti() in src/cti.cpp)
image_out_A, image_out_B, image_out_C, image_out_D = [
    arctic.remove_cti(
           image=image,
           n_iterations=5,
           parallel_roe=roe,
           parallel_ccd=ccd,
           parallel_traps=traps,
           parallel_express=5,
           verbosity=1,
    )
    for image in [image_A, image_B, image_C, image_D]
]

# Save the corrected image
autocti.acs.output_quadrants_to_fits(
    file_path=data_path + "_out.fits",
    quadrant_a=image_out_A,
    quadrant_b=image_out_B,
    quadrant_c=image_out_C,
    quadrant_d=image_out_D,
    header_a=image_A.header,
    header_b=image_B.header,
    header_c=image_C.header,
    header_d=image_D.header,
    overwrite=True,
)

ArCTIc also incorporates a model of "pixel bounce", an effect of voltage lag during correlated double sampling, due to finite capacitance between the sample and reference (ground) voltages. Pixel bounce can create trails similar to serial CTI. It is implemented by defining something like pixel_bounce = cti.PixelBounce( kA=-0.1, kv=0, omega=10, gamma=0.9 ) then passing pixel_bounce=pixel_bounce as an extra/alternative variable to add_cti() or remove_cti() (there are also duplicate add_pixel_bounce() functions that add only pixel bounce, and not CTI. Pixel bounce exists only in the python wrapper, not the C++ core.

More examples adding or removing CTI trails from a test image are in the run_demo() function of test/test_arcticpy.py.

Run python3 test/test_arcticpy.py with -d or -b for demo or benchmark functions.

C++

ArCTIc can also be run directly as ./arctic with the following command-line options:

• -h, --help
  Print help information and exit.
• -v <int>, --verbosity=<int>
  The verbosity parameter to control the amount of printed information:
  • 0 No printing (except errors etc).
  • 1 Standard.
  • 2 Extra details.
• -d, --demo
  Execute the editable demo code in the run_demo() function at the very top of src/main.cpp. A good place to run your own quick tests or use arctic without any wrappers. The demo version adds then removes CTI from a test image.
• -b, --benchmark
  Execute the simple test run_benchmark() function in src/main.cpp, e.g. for profiling.

The C++ code can also be used as a library for other C++ programs. See the run_demo() function in src/main.cpp for examples adding and removing CTI trails from a test image, and the lib_test example described below.

Unit Tests

Tests are included for most individual parts of the code, organised with Catch2.

As well as making sure the code is working correctly, most tests are intended to be relatively reader-friendly examples to help show how all the pieces of the code work if you need to understand the internal details as a developer, alongside the more user-focused documentation.

Compile the tests with make test (or make all) in the top directory, then run with ./test_arctic.

Add arguments to select which tests to run by their names, e.g:

• *'these ones'* All tests that contain 'these ones'.
• ~*'not these'* All tests except those that contain 'not these'.
• -# [#filename] All tests in filename.cpp.

Compiling with make lib_test will create a simple example of using the shared object library (libarctic.so), which is run with ./lib_test.

A few python tests of the primary functions are included for the arcticpy wrapper. Compile the wrapper with make wrapper (or make all) in the top directory, then run with pytest test/test_arcticpy.py.

Files

A quick summary of the code files and their contents:

• makefile The makefile for compiling the code. See its header.
  • get_gsl.sh The script called by the makefile to install GSL.
• arctic, test_arctic The program and unit-test executables.
• libarctic.so The shared object library.
• src/ Source code files.
  • main.cpp
    Main program. See above and its documentation for the command-line options, and see run_demo() for an example of running user-editable code directly.
  • cti.cpp
    Contains the primary user-facing functions add_cti() and remove_cti(). These are wrappers for clock_charge_in_one_direction(), which contains the primary nested for loops over an image to add CTI to, in order: each column, each express pass (see below), and each row. (And then each step and each phase in the clocking sequence if doing multiphase clocking, see below.)
  • ccd.cpp
    Defines the CCD classes that describe how electrons fill the volume inside each (phase of a) pixel in a CCD detector.
  • roe.cpp
    Defines the ROE classes that describe the properties of readout electronics (ROE) used to operate a CCD detector.
  • traps.cpp
    Defines the Trap classes that describe the properties of a single species of charge traps.
  • trap_managers.cpp
    Defines the internal TrapManager classes that organise one or many species of traps. Contains the core function n_electrons_released_and_captured(), called by clock_charge_in_one_direction() to model the capture and release of electrons and track the trapped electrons using the "watermarks".
  • util.cpp
    Miscellaneous internal utilities.
• include/ The *.hpp header files for each source code file.
• test/ Unit tests and examples.
• build/ Compiled object and dependency files.
• arcticpy/ The python, Cython, and other files for the wrapper.
  • setup.py The file for compiling the package.
  • src/ Source files.
    • cti.py
      The python versions of the primary user-facing functions add_cti() and remove_cti().
    • ccd.py, roe.py, traps.py
      The python versions of the CCD, ROE, and Trap classes that are needed as arguments for the primary CTI functions. These mirror the inputs for the corresponding same-name C++ classes documented below.
    • pixel_bounce.py
      Definition of the PixelBounce class, plus user-facing functions add_pixel_bounce() nd remove_pixel_bounce().
    • wrapper.pyx
      The Cython wrapper that passes python inputs to the C++ interface.
    • interface.cpp, interface.hpp
      The source and header files for functions to cleanly interface between Cython and the main precompiled library. e.g. converts the image array and CTI model inputs into the required C++ objects.
    • wrapper.cpp, ../wrapper.cpython*.so
      Compiled Cython output files.

Documentation

The code docstrings contain the full documentation for each class and function.

Most of the python wrapper code precisely mirrors the core C++ classes and functions. The full docstrings are not duplicated in that case so please refer to the C++ docstrings for the complete details.

This section provides an overview of the key contents and features. It is aimed at general users plus a few extra details for anyone wanting to navigate or work on the code itself.

The primary functions to add and remove CTI take as arguments custom objects that describe the trap species, CCD properties, and ROE details (see below). A core aspect of the code is that it includes several polymorphic versions of each of these classes. These provide a variety of ways to model CTI, such as different types of trap species, multiple phases in each pixel, or alternative readout sequences for trap pumping, etc.

Add/remove CTI

Add CTI

To add (and remove) CTI trails, the primary inputs are the initial image followed by the properties of the CCD, readout electronics (ROE), and trap species, for either or both parallel and serial clocking.

These parameters are set using the CCD, ROE, and Trap classes, as described below.

See add_cti()'s docstring in cti.cpp for the full details, and clock_charge_in_one_direction() for the inner code that loops over the image.

Remove CTI

Removing CTI trails is done by iteratively modelling the addition of CTI, as described in Massey et al. (2010) section 3.2 and Table 1.

The remove_cti() function takes all the same parameters as add_cti() plus the number of iterations for the forward modelling.

More iterations provide higher accuracy at the cost of longer runtime. In practice, 2 or 3 iterations are usually sufficient.

Image

The input image should be a 2D array of charge values, where the first dimension runs over the rows of pixels and the second inner dimension runs over the separate columns, as in this example of an image before and after calling add_cti() (with arbitrary trap parameters):

// Initial image with one bright pixel in the first three columns:
{{   0.0,     0.0,     0.0,     0.0  },
 { 200.0,     0.0,     0.0,     0.0  },
 {   0.0,   200.0,     0.0,     0.0  },
 {   0.0,     0.0,   200.0,     0.0  },
 {   0.0,     0.0,     0.0,     0.0  },
 {   0.0,     0.0,     0.0,     0.0  }}
// Image with parallel and serial CTI trails:
{{   0.00,    0.00,    0.00,    0.00 },
 { 194.06,    0.98,    0.49,    0.25 },
 {   1.96,  190.22,    1.92,    0.97 },
 {   0.99,    2.89,  186.47,    2.82 },
 {   0.50,    1.46,    3.80,    0.06 },
 {   0.25,    0.74,    1.92,    0.03 }}
// Image after correction for trailing:
{{   0.00,    0.00,    0.00,    0.00 },
 { 200.00, 1.30e-4, 3.15e-5,    0.00 }
 {2.05e-4, 199.999, 8.88e-4, 2.56e-4 }
 {1.93e-5, 1.20e-2, 199.994, 3.11e-3 }
 {   0.00, 2.53e-4, 3.95e-3,    0.00 }
 {   0.00,    0.00, 1.03e-3,    0.00 }}

As this illustrates, by default, charge is transferred "up" from row N to row 0 along each independent column, such that the charge in the first element/pixel 0 undergoes 1 transfer, and the final row N is furthest from the readout register so undergoes N+1 transfers. The CTI trails appear behind bright pixels as the traps capture electrons from their original pixels and release them at a later time.

Parallel clocking is the transfer along each independent column, while serial clocking is across the columns and is performed after parallel clocking, if the arguments for each are not omitted.

Note that technically instead of actually moving the charges past the traps in each pixel, as happens in the real hardware, the code tracks the occupancies of the traps (see Watermarks below) and updates them by scanning over each pixel. This simplifies the code structure and keeps the image array conveniently static.

Speedup 1: Express

As described in more detail in Massey et al. (2014) section 2.1.5, the effects of each individual pixel-to-pixel transfer can be very similar, so multiple transfers can be computed at once for efficiency.

This allows much faster computation with a mild decrease in accuracy.

For example, the electrons in the pixel closest to the readout have only one transfer, those in the 2nd pixel undergo 2, those in the 3rd have 3, and so on. The express input sets the number of times the transfers are calculated. express = 1 is the fastest and least accurate, express = 2 means the transfers are re-computed half-way through the readout, up to express = N where N is the total number of pixels, for the full computation of every step without assumptions.

The default express = 0 is a convenient input for automatic express = N.

Note that the total charge in an image is guaranteed to be conserved only with express = 0 (and also empty_traps_for_first_transfers = True if the trail length is comparable to the image size).

Speedup 2: Watermark pruning

With large, noiseless images in particular (and especially with slow capture traps), it is possible to accumulate a large number of watermarks containing negligible numbers of electrons. These increase runtime without affecting output. Packets of fewer than [parallel/serial]_prune_n_electrons can be moved into neighbouring watermarks every [parallel/serial]_prune_frequency readout steps. Default values are 1e-181 and 20`, but significant speedups are possible by tuning these for different images and different species of charge trap.

Offsets and windows

It is possible to (more quickly) process part of an image in two ways. In either use, because of edge effects, the region of interest should be expanded to include several pixels closer to readout (when adding CTI) or in all directions (when correcting CTI).

Either pass the full image, using the window_start and _stop arguments to indicate the first and last pixel numbers to be processed; or pass a subset of the image and use offset to indicate the number of missing, preceding pixels.

Partial readout

TBD

CCD

How electrons fill the volume inside each (phase of a) pixel in the charged-coupled device (CCD) detector.

By default, charge is assumed to move instantly from one pixel to another, but each pixel can be separated into multiple phases, in combination with a multiphase ROE clock sequence.

See the CCD and CCDPhase class docstrings in ccd.cpp for the full documentation.

Multiple phases

The CCD object can be created either with a single CCDPhase or a list of phases plus an array of the fraction of traps distributed in each phase, which could be interpreted as the physical width of each phase in the pixel.

The number of phases must be the same for the CCD and ROE objects.

ROE

The properties of readout electronics (ROE) used to operate a CCD.

Three different modes are available:

• Standard, in which charge is read out from the pixels in which they start to the readout register, so are transferred across a different number of pixels depending on their initial distance from readout.
• Charge injection, in which the electrons are directly created at the far end of the CCD, then are all transferred the same number of times through the full image of pixels to the readout register.
• Trap pumping (AKA pocket pumping), in which charge is transferred back and forth, to end up in the same place it began.

See the ROE, set_clock_sequence(), and child class docstrings in roe.cpp for the full documentation, including illustrative diagrams of the multiphase clocking sequences.

Pre-scan and over-scan

Use prescan_offset to specify the number of physical prescan pixels that are present in the hardware but always absent from stored data arrays. This works in exactly the same way (and adds to) any window_offset.

Use overscan_start to specify the first pixel in a supplied data array that is virtual overscan. This effectively defines the number of physical pixels in the CCD as overscan_start-1. Unfortunately, this needs to be specified here rather than in the CCD structure.

Express matrix

The ROE class also contains the set_express_matrix_from_pixels_and_express() function used to generate the array of express multipliers that controls which transfers are computed.

Multiple phases

Like the CCD, the ROE object can model single or multiple steps in the clock sequence for each transfer. The number of steps in a clocking sequence is usually same as the number of phases, but not necessarily, as in the case for trap pumping.

The number of phases must be the same for the CCD and ROE objects.

Trap species

The parameters for a trap species.

See the Trap* class docstrings in traps.cpp for the full documentation.

Instant capture

For the relatively simple algorithm of release first then instant capture. This is the primary model used by previous versions of ArCTIC.

Optionally, these traps can be assigned a non-uniform distribution with volume within the pixel, e.g. to model "surface" traps that are only reached by very large charge clouds.

Slow capture

For combined release and non-instant capture, following Lindegren (1998), section 3.2.

Continuum lifetime distribution (instant capture)

For a trap species with a continuum (log-normal distribution) of release timescales, and instant capture.

Continuum lifetime distribution (slow capture)

For a trap species with a continuum (log-normal distribution) of release timescales, and non-instant capture.

Trap managers

This is not relevant for typical users, but is key to the internal structure of the code and the "watermark" approach to tracking the trap occupancies (see below).

The different trap manager child classes also implement the different algorithms required for the corresponding types of trap species described above, primarily in the n_electrons_released_and_captured() method.

See the TrapManager* class docstrings in trap_managers.cpp for the full documentation.

To allow the options of multiple types of trap species and/or multiple phases in each pixel, the code actually uses a top-level "trap-manager manager" to hold the necessary multiple TrapManager objects. See the TrapManagerManager class docstring for the details.

Watermarks

The watermark_volumes and watermark_fills arrays track the states of the charge traps and the number of electrons they have captured. The core release and capture algorithms are based on these arrays.

watermark_volumes is a 1D array of the fractional volume each watermark represents as a proportion of the pixel volume. These are set by the volume the charge cloud reaches in the pixel when electrons are captured.

watermark_fills is a 2D-style array (stored as 1D internally) of the fraction of traps within that watermark that are full, i.e. that have captured electrons, for each trap species. For efficiency, these values are multiplied by the density of that trap species (the number of traps per pixel).

In the standard case, capture of electrons creates a new watermark level at the "height" of the charge cloud and can overwrite lower ones as the traps are filled up, while release lowers the fraction of filled traps in each watermark level, without changing the volumes. Note that the stored volumes give the size of that level, not the cumulative total.

The details can vary for different trap managers.

The unit tests in test/test_trap_managers.cpp, especially those for release and capture, contain simple-number examples to demonstrate how it all works for developers.

Python wrapper

After compiling the Cython, the arcticpy python module can be imported and used as normal. test/test_arcticpy.py contains some tests and a basic example, and see the full example at the top of this file.

The majority of the python functions and classes (arcticpy/src/*.py) directly mirror the core C++, so in those cases the full docstrings are not duplicated.

The wrapper is organised internally as follows:
python --> Cython --> C++ wrapper --> core library.
This multi-level structure is a bit more extensive then strictly necessary, but this keeps each level much cleaner and with a single purpose.

The user-facing python functions take numpy arrays and custom input-parameter objects as user-friendly arguments. These mirror exactly the custom C++ objects used as arguments by the core C++ program described above. To convert cleanly between the two, the individual arrays and numbers are extracted from the python objects and are passed via the Cython wrapper to the C++ wrapper, which then builds the C++ objects as arguments for the core library functions.

Version history

• v7 (2022, C++/python) Translation of v6, now back to full speed. Includes all features seen in Euclid CCDs before launch.

• v6 (2020, cython/python) Jacob Kegerreis implements non-instantaneous charge capture, distribution of charge release times within each species, non-uniform spatial distribution of e.g. surface traps, sophisticated readout for inter-pixel traps, charge injection, or trap pumping. Much slower than v5.

• v5 (2015, C++) Adaptive 'neo2' gridding of traps by splitting the continuous field only at each electron fill levels, and recombining grid cells when traps refill at new high watermark (Massey et al. 2015)

• v4 (2014, C++) Oliver Cordes and Ole Marggraf implement huge speed up. Monitors the high water mark of signal electrons, and only considers traps that could have been filled. Post-correction noise-whitening. (Massey et al. 2014).

• v3 (2010, IDL) Richard Massey implements gradual tradeoff between accuracy and speed, through variable EXPRESS option. Inter-pixel traps confirmed to be degenerate with change of effective density, and release profile well-fit by sum of exponentials. Hubble Space Telescope model updated following shuttle servicing mission (Massey 2010).

• v2 (2009, Java/IDL) Assumes a fixed grid of fractional traps (and introduces concept of well fill level) to reduce noise. Charge trap parameters measured from hot/warm pixels in Hubble Space Telescope imaging (Massey et al. 2009). Later converted to python by STScI, with EXPRESS=1 speedup also used by and empirical f(t) trap release profile. Capture confirmed empirically to be instant.

• v1 (2008, Java) Chris Stoughton extends Fortran77 code by Bristow (2003), introducing 3D pixel structure, multiple trap species, and reducing runtime by moving traps not charge. Discrete traps are distributed at random, which adds noise, and are monitored during every transfer, which is slow. Predicted effect for SNAP telescope (Rhodes et al. 2010).

Older algorithms for CTI correction either approximated trailing as convolution with a flux-dependent kernel (e.g. Rhodes et al. 2000) or were additive/multiplicative factors applied to object flux/position/shape/etc at a catalogue level (e.g. Riess et al. 2000, 2003, Rhodes et al. 2007).