Devsim

Explicit math library loading

Introduction

Since the Intel Math Kernel Library started versioning the names of their dynamic link libraries, it has been difficult to maintain a proper Anaconda Python environment when the version has been updated. With this release, it is possible to use any recent version of the Intel MKL. In addition, the user is able to load alternative BLAS/LAPACK math libraries.

Intel MKL

From DEVSIM Version 2.1.0 onward, a specific version is not required when loading the Intel MKL. If the Intel MKL is not found, the import of the ``devsim`` module will fail, and an error message will be printed. This method is the default, and should work when using an Anaconda Python environment with the ``mkl`` package installed.

When using a different Python distribution, or having an installation in a different place, it is possible to specify the location by modifying the ``LD_LIBRARY_PATH`` environment variable on Linux, or using ``DYLD_LIBRARY_PATH`` on macOS. The explicit path may be set to the MKL math libraries may be set using the method in the next section.

Loading other math libraries

It is possible to load alternative implementations of the BLAS/LAPACK used by the software. The ``DEVSIM_MATH_LIBS`` environment variable may be used to set a ``:`` separated list of libraries. These names may be based on relative or absolute paths. The program will load the libraries in order, and stop when all of the necessary math symbols are supplied. If symbols for the Intel MKL are detected, then the Pardiso direct solver will be enabled.

New CMAKE build option
For those building the software, the ``EXPLICIT_MATH_LOAD`` CMAKE option has been added to control the new explicit math loading feature. An important benefit of this option is that it is possible to build a release version of the software, even if the Intel MKL has not been installed on the build computer.

Direct solver selection

The direct solver may be selected by using the ``direct_solver`` parameter.

devsim.set_parameter(name='direct_solver', value='mkl_pardiso')


The following options are available:
- ``mkl_pardiso`` Intel MKL Pardiso
- ``superlu`` SuperLU 4.3

The default is ``mkl_pardiso`` when the Intel MKL is loaded. Otherwise, the default will switch to ``superlu``.

Kahan summation in extended precision mode

The ``kahan3`` and ``kahan4`` functions are now using the Kahan summation algorithm for extended precision model evaluation. Previously, this algorithm was replaced with 128-bit floating point addition and subtraction in releases that support extended precision mode. With this change, better than 128-bit floating precision is available when extended precision is enabled.

devsim.set_parameter(name = "extended_model", value=True)


The ``testing/kahan_float128.py`` test has been added.

Visual Studio 2022

The Microsoft Windows``win64`` release version is now built using the Visual Studio 2022 compiler. For users needing extended precision on the Windows platform, the ``msys`` build is recommended.

Update documentation files

The following files were updated in the text documentation distributed with the software.
- ``CONTRIBUTING.md``
- ``INSTALL.md``
- ``README.md``

This was done to create a version to coincide with this paper in the Journal of Open Source Software.

Sanchez, J. E., (2022). DEVSIM: A TCAD Semiconductor Device Simulator. Journal of Open Source Software, 7(70), 3898, [https://doi.org/10.21105/joss.03898](https://doi.org/10.21105/joss.03898)

Update MKL Version

The release version of this software is build against version 2 of the Intel MKL, which corresponds to the latest version of Anaconda Python. If you have issues running ``DEVSIM`` with this new version, please contact us at [https://forum.devsim.org](https://forum.devsim.org) for assistance.

Update SYMDIFF reference

The submodule reference to ``SYMDIFF`` was updated due some changes to its build files. Some ``SYMDIFF`` build scripts in the ``DEVSIM`` repository were also updated.

Versioned MKL DLL in release build

The Intel Math Kernel Library now uses versioned library names. Binary releases are now updated against the latest versioned dll names from MKL available in the Anaconda Python distribution.

Fixed issue in ramp function.

The ``rampbias`` function in the ``devsim.python_packages.ramp`` module has been fixed to properly reduce the bias when there is a convergence failure.

Transient Simulation

Fixed bug with ``transient_tr`` (trapezoidal) time integration method where the wrong sign was used to integrate previous time steps.

Fixed bug in the charge error calculation, which calculates the simulation result with that a forward difference projection.

Added ``testing/transient_rc.py`` test which compares simulation with analytic result for RC circuit.

Added ``set_initial_condition`` command to provide initial transient conditions based on current solution.

Create interface from node pairs

Added ``create_interface_from_nodes`` to make it possible to add interface from non-coincident pairs of nodes.

Solver

Convergence Tests

The ``maximum_error`` and ``maximum_divergence`` options where added to the ``solve`` command. If the absolute error of any iteration goes above ``maximum_error``, the simulation stops with a convergence failure. The ``maximum_divergence`` is the maximum number of iterations that the simulator error may increase before stopping.

Verbosity

During the ``solve``, circuit node and circuit solution information is no longer printed to the screen for the default verbosity level. In addition, the number of equations per device and region is no longer displayed at the start of the first iteration.

SuperLU

The code now supports newer versions of ``SuperLU``. The release version is still using SuperLU 4.3 for the iterative solution method, and the Intel MKL Pardiso for the direct solve method.

Simulation Matrix

The ``get_matrix_and_rhs`` command was not properly accepting the ``format`` parameter, and was always returning the same type.

Build Scripts

The build scripts have been updated on all platforms to be less dependent on specific Python 3 versions.

An updated fedora build script has been added. It uses the system installed ``SuperLU`` as the direct solver.

Documentation Files

Some out of date files (e.g. RELEASE, INSTALL, . . .) have been removed. The [README.md](README.md) has been updated and the [INSTALL.md](INSTALL.md) have been updated.

Command Options

The ``variable_name`` option is no longer recognized for the ``devsim.contact_equation`` and ``devsim.interface_equation`` as it was not being used.

Array Type Input and Output

In most circumstances, the software now returns numerical data using the Python ``array`` class. This is more efficient than using standard lists, as it encapsulates a contiguous block of memory. More information about this class can be found at [https://docs.python.org/3/library/array.html](https://docs.python.org/3/library/array.html). The representation can be easily converted to lists and ``numpy`` arrays for efficient manipulation.

When accepting user input involving lists of homogenous data, such as ``set_node_values`` the user may enter data using either a list, string of bytes, or the ``array`` class. It may also be used to input ``numpy`` arrays or any other class with a ``tobytes`` method.

Get Matrix and RHS for External Use

The ``get_matrix_and_rhs`` command has been added to assemble the static and dynamic matrices, as well as their right hand sides, based on the current state of the device being simulated. The ``format`` option is used to specify the sparse matrix format, which may be either in the compressed column or compressed row formats, ``csc`` or ``csr``.

Maximum Divergence Count

If the Newton iteration errors keep increasing for 20 iterations in a row, then the simulator stops. This limit was previously 5.

Mesh Visualization Element Orientation

Elements written to the ``tecplot`` format in 2d and 3d have node orderings compatible with the element connectivity in visualization formats. Specifying the ``reorder=True`` option in ``get_element_node_list`` will result in node ordering compatible with meshing and visualization software.

Math Functions

The following inverse functions and their derivatives are now available in the model interpreter.
- ``erf_inv`` Inverse Error Function
- ``erfc_inv`` Inverse Complimentary Error Function
- ``derf_invdx`` Derivative of Inverse Error Function
- ``derfc_invdx`` Derivative of Complimentary Inverse Error Function

The Gauss-Fermi Integral, using Paasch's equations are now implemented.
- ``gfi`` Gauss-Fermi Integral
- ``dgfidx`` Derivative of Gauss-Fermi Integral
- ``igfi`` Inverse Gauss-Fermi Integral
- ``digfidx`` Derivative of Inverse Gauss-Fermi Integral

Each of these functions take two arguments, ``zeta`` and ``s``. The derivatives with respect to the first argument are provided. Please see ``testing/GaussFermi.py`` for an example.

In extended precision mode, the following functions are now evaluated with full extended precision.
- ``Fermi``
- ``dFermidx``
- ``InvFermi``
- ``dInvFermidx``

The following double precision tests:

- ``testing/Fermi1.py`` Fermi Integral Test
- ``testing/GaussFermi.py`` Gauss Fermi Integral Test

Have extended precision variants:

- ``testing/Fermi1_float128.py``
- ``testing/GaussFermi_float128.py``

Installation Script

A new installation script is in the base directory of the package.
It provides instructions of completing the installation to the ``python`` environment without having to set the ``PYTHONPATH`` environment variable.
It notifies the user of missing components to finish the installation within an ``Anaconda`` or ``Miniconda`` environment.


To use the script, use the following command inside of the ``devsim`` directory.


python install.py


The install script will write a file named ``lib/setup.py``, which can be used to complete the installation using ``pip``. The script provides instructions for the installation and deinstallation of ``devsim``.


INFO: Writing setup.py
INFO:
INFO: Please type the following command to install devsim:
INFO: pip install -e lib
INFO:
INFO: To remove the file, type:
INFO: pip uninstall devsim

The ``custom_equation`` command has been modified to require a third return value. This boolean value denotes whether the matrix entries should be row permutated or not. For the bulk equations this value should be ``True``. For interface and contact boundary conditions, this value should be ``False``.

It is now possible to replace an existing ``custom_equation``.

The file ``examples/diode/diode_1d_custom.py`` demonstrates custom matrix assembly and can be directly compared to ``examples/diode/diode_1d.py``.

The ``EdgeNodeVolume`` model is now available for the volume contained by an edge.

The ``contact_equation`` command now accepts 3 additional arguments.

- ``edge_volume_model``
- ``volume_node0_model``
- ``volume_node1_model``

These options provide the ability to do volume integration on contact nodes.

The ``equation`` command has replaced the ``volume_model`` option with:

- ``volume_node0_model``
- ``volume_node1_model``

so that nodal quantities can be more localized.

More details are in the manual.