New block : NetlistStateSpace

[Pimss]

Summary

This PR adds support for importing linear SPICE-like netlists into PathSim as state-space models.

The new NetlistStateSpace block converts RLC circuits and independent sources into an explicit state-space representation using Modified Nodal Analysis (MNA), allowing circuits described by netlists to be directly integrated into PathSim simulations.

Main changes

• Added netlist_to_statespace.py with:

  • SPICE-like netlist parsing
  • SI unit and .param handling
  • R, L, C, V and I elements
  • coupled inductors through K elements
  • behavioral current and voltage source support
  • Modified Nodal Analysis formulation
  • algebraic-variable elimination
  • explicit state-space generation
  • symbolic and fast reduction modes
  • configurable voltage and current outputs
  • inline netlist string support
  • netlist file support

• Added NetlistStateSpace to the public pathsim_rf API.

• Added sympy as a dependency for symbolic reduction.

• Added python-mumps as a dependency for fast numerical reduction.

• Added parser tests covering:

  • SI unit parsing
  • parameter substitution
  • behavioral sources
  • invalid elements
  • invalid outputs
  • current-source handling
  • MUMPS fallback
  • voltage and current output selection

• Added state-space tests covering:

  • netlist file input
  • inline netlist input
  • feedback current-source inputs
  • voltage and current outputs
  • symbolic reduction
  • fast numerical reduction
  • fast/symbolic equivalence
  • invalid inputs

• Added validation netlists:

  • tests/filter.net
  • tests/filter_without_load.net

• Added the netlist_import.ipynb documentation example demonstrating:

  • netlist import
  • PathSim state-space generation
  • comparison against a manually constructed PathSim model
  • behavioral load feedback

Files added

• src/pathsim_rf/netlist_to_statespace.py
• tests/test_netlist_parser.py
• tests/test_netlist_statespace.py
• tests/filter.net
• tests/filter_without_load.net
• docs/source/examples/netlist_import.ipynb

Files modified

• src/pathsim_rf/__init__.py
• pyproject.toml

Validation

The new test suite validates the netlist parser and the generated state-space models. Both symbolic and fast reduction paths are tested, including their numerical equivalence.

The added documentation example provides an end-to-end validation of a netlist-derived state-space model against an equivalent PathSim implementation.

[Pimss]

Small problem, python_mumps is only compatible with python>=3.11 and pathsim_rf has release starting from 3.9.
Currently looking into it

[milanofthe]

Do we really need mumps? Wouldnt scipy superLU do for the sparse solve?

[Pimss]

This would be possible. However, on large netlist, I suspect it will be significantly slower. I am currently checking if python-mumps 0.0.3 (compatible with python 3.9) is going to work with current implementation

[Pimss]

0.0.3 is compatible with python 3.9 but there may be some rework to do. I intend to make python-mumps only available starting python 3.11. The code already have a fallback to sympy if python-mumps is not installed so this would work with python<=3.10 (sympy), python>= 3.11 (sympy and/or python-mumps)

[Pimss]

There may be some other issues. python-mumps package from pypi do not include mumps... (conda/mamba package does)
env install may works, but tests will most likely fail.

[Pimss]

note : the addition of K coupling can make the system not that sparse. on a netlist with 396 k, mumps implementation takes 0.8s, scipy more than 10 minutes.

[Pimss]

Current status : passed sympy LU to scipy LU solve.

On a netlist with 31 L, 64 R and 397 k
sympy LU >=10 minutes (not finished - removed)
sympy + scipy LU = 21 s (latest commit)
mumps =0.58 s (latest commit)

python-mumps is installed with mamba and pypi have no way of installing mumps (pypi python-mumps install does not have mumps included)

I see two options:
- Leave the code as it is, remove python-mumps from pyproject.toml and leave the user install python-mumps (and pathsim) with a conda env
- Remove all python-mumps implementation to simplify the code

I recommend solution 1, test including python-mumps will be skipped (TestFastReduction class in tests/test_netlist_statespace.py)

[milanofthe]

I will check it. Factor of almost 40 between scipy superLU and mumps seems insane. Also a linear system of n roughly 100 should solve way faster then that.

Generally I am for less dependencies.

[Pimss]

Here is the use case I'm using MediumCoupled.txt
Mumps is very efficent. Maybe the factor 40 is a bit much, however I used it on systems with more that 16M k (in another implementation) and mumps is impressive.

Nevertheless the argument of less dependencies is perfectly understandable.

from pathsim_rf.netlist_to_statespace import NetlistStateSpace
state_space_sym = NetlistStateSpace("MediumCoupled.txt", output_voltages=["N3"], reduction_mode="symbolic")
state_space_mumps = NetlistStateSpace("MediumCoupled.txt", output_voltages=["N3"], reduction_mode="fast")

[milanofthe]

Thanks for the netlist. The time goes into sp.nsimplify while stamping (18.9 s); float64 stamping with scipy splu gives the same model in 6 ms. Please stamp with numpy and reduce with scipy only, and drop sympy and python-mumps.

Second issue: this netlist has inductor cutsets (no capacitors, 22 node groups reach ground only through inductors). G_alg_a is singular and gets silently regularized, which adds 22 eigenvalues around 1e17 to A. That needs a proper reduction or a clear error.

[Pimss]

Yes, my current tests are consistent with this

[Pimss]

Current implementation remove mumps and sympy dependencies.
Tests, doc and env updated.
This version was tested on a netlist with 738718 k, 1215 R and 1215 L.
Scipy performance is not a bottleneck (around 7 sec for this use case), so mumps is not mandatory
The eigenvalues issue is not resolved yet.

[milanofthe]

Is this because the circuits are degenerate and you have cutsets of inductors etc?

[Pimss]

The matrix E can be singular in my use cases. Matrix L is dense due to k couplings. I expect to code to work for classic circuit topologies but this may be a problem for heavy coupled problem like the few I'm using.
I have a plan, I already encounter this issue, and some friends helped. They gave me some advice on how to proceed.
So I will try a new implementation using MassMatrixDAE

[Pimss]

New implementation

from mass_matrix_netlist import MMDAENetlist
test_mmn = MMDAENetlist("filter.net")

This expose the full problem to pathsim so wiring is more convenient with labels

conn_0 = Connection(pulsesource[0], test_mmn["V1"])
...
conn_14 = Connection(test_mmn["V(n1)"],  scope[5])

I believe Pathview only use index so this will be a problem for large netlist.
Despite been ordered in the following order : L currents, R currents, C currents, voltage node and Source currents, indexes will be confusing for any gui user.

[milanofthe]

PathView can use the port labels of the block. So if there is a way that you construct them from the netlist (net labels, or even from the parameters) this should work.