Accelerated electrical, thermal, EMC, and coupled PCB analysis

Status: beta. pcb-analysis is under active development. The public API, the JSON schemas and the defaults can change between minor releases (0.9.0 renamed the current-field contract), and the numerical results have been validated on a small set of boards so far. Please report problems and mismatches as GitHub issues.

Accelerated PEEC and matrix-free FEM solvers for electrical and thermal PCB analysis, radiated-emission evaluation for EMC, and PDN optimization.

The repository has three solver families: electrical, thermal and emc.

These solvers are designed to analyse PCBs faster, so that they can be embedded in an automated, LLM-driven PCB design workflow.

Features

• electrical: 2.5D sheet PEEC for current density, voltage drop and skin effect on multilayer copper; matrix-free FEM for DC conduction and 2D frequency-domain Maxwell; 3D voxel PEEC (PyPEEC) for thick conductors; DICE delta scoring to rank many candidate edits without a full solve.
• thermal: steady and transient heat conduction through the board stack, with convection, radiation and fixed-temperature boundaries.
• emc: near-field scans, far-field patterns and CISPR 32 / FCC Part 15 margins computed from a solved current distribution.
• multiphysics: electro-thermal fixed point with temperature-dependent copper, and emission of the cold or heated current.
• geometry: reads a KiCad (or any mechanical) STEP export into the solvers' grids, keeping the plated via and through-hole barrels.
• NumPy on the CPU everywhere; CUDA (CuPy) and OpenMP C++ kernels where they pay off.

Install

pip install pcb-analysis            # CPU only
pip install 'pcb-analysis[cuda]'    # adds CuPy for the CUDA paths

Requirements:

• Python 3.11 or later. The base install includes OpenCASCADE (cadquery-ocp) and PyPEEC.
• Optional: an NVIDIA driver with CUDA 13.x for [cuda]. Run nvidia-smi to check it.
• kicad-cli (KiCad 8 or later) for the KiCad path below.

Each release is also attached to its GitHub release (wheel, sdist and SHA-256 checksums). To install one without PyPI:

pip install 'git+https://github.com/AFLOY/pcb-analysis.git@v0.9.0'

How to use: start from KiCad

KiCad is the recommended environment. Design the board there, export it as STEP with its copper, and let pcb-analysis read the copper back onto a grid.

1. Export the board with its copper. --include-tracks, --include-pads and --include-zones write the copper as solids. Vias and through-holes then come out as the plated barrels KiCad models, and pcb-analysis reads those barrels as they are, so do not add --fill-all-vias.

kicad-cli pcb export step --include-tracks --include-pads --include-zones \
  --include-inner-copper --no-extra-pad-thickness --no-components --force \
  --output power_module.step power_module.kicad_pcb

2. Solve it. The stackup comes from the .kicad_pcb. The body map sorts the solids into the board, one copper sheet per layer and the barrels. Two pads act as the input faces: 1 A enters at the input connector's VIN pin (J1 pad 3) and leaves at the MOSFET's source pad (Q1 pad 3). The API has no pad object: each pad's centre is read from the KiCad board (the pad properties, in mm), and the copper cells under it become the terminal's cells. current_field_problem_mapping collects the grid, copper, barrels and terminals into one current-field problem.

import numpy as np
from geometry.cad_import import (
    board_barrels, board_vias, kicad_step_body_map, layers_from_kicad_stackup,
    load_step, current_field_problem_mapping, rasterize_board, read_kicad_stackup, resolve_bodies,
)
from electrical.sheet_peec import solve_current_field_problem

board = "power_module.kicad_pcb"
layers, board_top_mm = layers_from_kicad_stackup(read_kicad_stackup(board))
body_map = kicad_step_body_map(layers, board_top_z_mm=board_top_mm)   # board, copper layers, barrels
resolved = resolve_bodies(load_step("power_module.step"), body_map)
raster = rasterize_board(resolved, body_map.board, pitch_mm=0.25, y_down=True)

def pad_cells(layer, x_mm, y_mm, half_width_mm=0.5):
    """Copper cells of `layer` around a pad centre given in KiCad coordinates."""
    row, col = raster.cell_of(x_mm * 1e-3, -y_mm * 1e-3)   # the STEP export negates KiCad's y
    k = [spec.name for spec in raster.layers].index(layer)
    r = int(np.ceil(half_width_mm / raster.pitch_mm))
    rows, cols = raster.shape
    return [{"layer": layer, "x": c, "y": y}
            for y in range(max(row - r, 0), min(row + r + 1, rows))
            for c in range(max(col - r, 0), min(col + r + 1, cols)) if raster.occupancy[k, y, c] > 0]

terminals = [
    {"name": "VIN", "pad": "J1.3", "current_a": 1.0, "cells": pad_cells("F.Cu", 129.0, 97.58)},
    {"name": "SRC", "pad": "Q1.3", "current_a": -1.0, "cells": pad_cells("F.Cu", 149.26, 95.675)},
]
problem = current_field_problem_mapping(raster, terminals=terminals, frequency_hz=0.0,
                                    vias=board_vias(resolved, raster), barrels=board_barrels(resolved, raster))
result = solve_current_field_problem(problem)             # DC sheet PEEC on the CPU
print(f"voltage span {result.metrics['voltage_span_v'] * 1e3:.3f} mV, "
      f"peak current density {result.metrics['max_current_density_a_per_mm2']:.2f} A/mm2")

On the example board (two layers, a 139 × 139 grid) this takes a few seconds on a desktop CPU.

Next steps

Each package README continues from the same export, one step at a time:

Documentation

The methods, their limits and the measurements behind each design decision:

• Architecture and design and requirements
• Sheet PEEC, matrix-free MPIR-FEM, 3D voxel PEEC
• Thermal MPIR-FEM
• EMC dipole superposition
• Coupled scenarios
• STEP and KiCad input
• Reports: sheet PEEC versus matrix-free FEM, Maxwell accuracy and CPU timing

Release and migration notes are in CHANGELOG.md.

Development

pip install -e '.[test]'              # from a clone; CPU only, as CI runs it
python -m pytest tests/
pip install -e '.[native]'            # optional C++ kernels (pybind11, OpenMP)
cmake -S . -B build/native -DCMAKE_BUILD_TYPE=Release
cmake --build build/native

Tests that need the C++ kernels are skipped until they are built. CMakeLists.txt lists the native targets; -DPCB_NATIVE_OPENMP=OFF drops OpenMP and -DPCB_NATIVE_MARCH=x86-64-v3 (for example) replaces -march=native. The scripts in experiments/ regenerate the numbers quoted in docs/; run them from the repository root.

License

MIT