A 10 kW, 800 V to 400 V, 100 kHz dual-active-bridge designed through the AIPE workflow engine: requirements → analytical single-phase-shift design → executed ngspice switching simulation → waveform-driven magnetics, device and thermal design → Gmsh/Elmer field checks and FreeCAD geometry → Core Engineering State with evidence → review package. Fourteen tasks across eight specialist repositories in one checkpointed graph. Maturity: prototype (software, circuit-simulation and field-simulation gates passed; the CAD/PCB gate executed but fails on unfinished board routing; measurement and hardware qualification not claimed).
Exact lossless SPS model in AIPE Design Agent (pea.converters.dab_sps): 22.5° phase shift transfers exactly 10 kW with 35 µH and a 2:1 turns ratio; 13.68 A RMS series current; ZVS current condition met on both bridges. The SPS digital control baseline (PI current loop at 2 kHz, 2P2Z voltage loop at 300 Hz, worst phase margin 55°) is designed on the cycle-averaged plant and its plant law is checked against executed ngspice runs within 0.07 %.
ngspice 47 runs the generated Level 1 (ideal) model: 9 996.8 W output, 13.678 A RMS series current, 22.5000° measured phase shift, periodicity error 2·10-6; time-step convergence from T/400 to T/1600 recorded. The Level 2 behavioural model (200 ns dead time, datasheet on-resistance, Coss, body-diode fit, DC-link parasitics, RC load) reaches about 98.6 % with every switch turning on under ZVS, and its HDF5 waveform dataset feeds every downstream stage.
AIPE-Magnetics-Database designs a TDK E 70/33/32 transformer (20:10, interleaved litz, 35 W, 1.55 µH leakage) from the executed currents; AIPE-Semiconductor-Database screens the SiC records on the per-switch stresses and evaluates Wolfspeed C4MS047120K / C3M0025065K at 9.0 W / 17.4 W per device from manufacturer PLECS data; AIPE Design Agent compares natural, forced-air and liquid cooling (63 °C worst junction) and screens plate-fin heat sinks on a seeded Pareto front.
AIPE-Field-Simulation runs Gmsh + Elmer on the selected core window (FEM leakage 1.48 µH, 4 % below the 1-D design value) and on the knee heat-sink candidate (hotspot 59.6 °C, 9 % above the correlation), each at two mesh levels with energy checks. AIPE-CAD-Agent builds the transformer and heat sink in FreeCAD from the same Engineering State with STEP export, mass accounting and interference and clearance checks.
Every task returns an execution envelope with tool versions, SHA-256 artifact hashes, model validity ranges and numerical checks; results enter the AIPE Core Engineering State as evidence (twelve preserved state versions, fourteen evidence records). aipe-workflows report writes JSON, Markdown and an AIPE-Doc styled HTML package with figures and gate statuses; aipe-workflows iterate records cross-module feedback loops with accept/reject decisions.
Windows or Linux workstation, Python 3.10+, sibling checkouts of AIPE-Core, AIPE-Registry, AIPE-Design-Agent, AIPE-Simulation-Skills, AIPE-Semiconductor-Database, AIPE-Magnetics-Database, AIPE-Field-Simulation, AIPE-CAD-Agent and AIPE-Workflows; ngspice, Gmsh, Elmer and FreeCAD from their official release channels (the doctor reports official download channels and licences for anything missing and installs nothing by itself; stages whose tool is absent are reported as blocked, never faked).
python -m pip install -e ../AIPE-Simulation-Skills -e ../AIPE-Design-Agent -e ../AIPE-Semiconductor-Database
python -m pip install -e ../AIPE-Magnetics-Database[hdf5] -e ../AIPE-Field-Simulation -e ../AIPE-CAD-Agent -e ../AIPE-Workflows
aipe-workflows doctor --qualify # detect and numerically qualify ngspice, KiCad, FreeCAD, Gmsh, Elmer
aipe-workflows plan examples/dab-10kw/specification.yaml --checkouts ..
aipe-workflows run examples/dab-10kw/specification.yaml --checkouts ..
aipe-workflows verify <run-id>
aipe-workflows report <run-id>In Claude Code or Codex, open AIPE-Workflows and ask: “Design and optimise a 10 kW DAB converter using AIPE, run the available simulations, and generate the engineering review package.” The agent discovers the workflow, qualifies the local tools and invokes the commands above.
Reference run of 21 October 2026 (ngspice 47, Gmsh 4.15.2, Elmer 9.0, FreeCAD 1.1.4, Windows 11); the full package with envelopes, netlists, summaries, FEM case files, CAD metadata and figures is in examples/dab-10kw/results.
| Stage | Executed result | Cross-check |
|---|---|---|
| Analytic SPS vs ngspice Level 1 | 10 000 W vs 9 996.8 W; 13.6775 A vs 13.6779 A RMS; 22.5000° measured | 5 W is the modelled resistive loss |
| Level 2 behavioural ngspice | ≈130 W loss, 810 V / 412 V peak device voltage, ZVS on all eight switches | capacitor energy drift corrected |
| Control baseline | PM 93.6° / 60.0° (current / voltage loop), worst 55.4° | plant law vs ngspice at 10°, 22.5°, 35°: max 0.07 % |
| Transformer design | E 70/33/32 N87, 20:10 litz P-S-P, 0.146 T, 35.0 W, 1.55 µH leakage | window FEM 1.48 µH (−4.4 %), two mesh levels |
| Device screening | C4MS047120K 9.0 W, C3M0025065K 17.4 W per device at 125 °C | manufacturer PLECS tensors (tier A); module CBB021M12FM3 ranks first overall |
| Cooling | liquid 63.0 °C, forced air 64.3 °C, natural 79.2 °C (secondary bridge, 4 devices per sink) | Stage A knee: Al 81 × 65 mm, 11 fins, 0.44 K/W, 0.28 kg |
| Heat-sink FEM / CAD | hotspot 59.6 °C, 0.562 K/W; CAD sink mass 0.275 kg | correlation 0.518 K/W (+8.6 %); mass −1.3 % |
Device parameters are manufacturer-specified datasheet or model values or labelled assumptions, not measurements; thermal and FEM film coefficients are correlation inputs, not CFD results; clearance requirements in the CAD checks are illustrative caller inputs.
The three KiCad boards generated by AIPE-PCB-Agent (primary power, secondary power, control) pass ERC, but their routing is unfinished: the workflow's ERC/DRC task reports 82 / 81 / 499 unconnected items and 22 / 2 / 0 DRC errors, so the cad_pcb gate reads fail for this run even though the FreeCAD checks passed. CFD, manufacturer SPICE models (Level 3), laboratory measurement and hardware qualification remain separate statuses that the software never asserts. The earlier analytical teaching series remains available: DAB circuit, modelling and control tutorial →