Skip to content

Iteration 3 — grounding rework, MISO topology change, MOSFET swap

Bench session at 13:09, 2026-03-05 — mid-iteration-3 testing
Bench session during iteration 3, March 5, 2026. The intermittent sensor-loss / status-corruption symptoms documented below were what triggered the iteration-4 grounding redesign.

Status: superseded. KiCad zip backups for iteration 3 are preserved at hardware/legacy/iteration-3/ — untrackedCHB_INVERTER.zip (the working tree at the iteration-3 freeze) and 2026-04-07_Full_Bridge_Backup.zip (a mid-iteration snapshot).

What was attempted

Iteration 3 committed to the per-bridge isolated supply plan that came out of iteration 2, and reworked the board to support it:

  • B0515S 5V→15V isolated DC-DC per bridge (one per side), feeding the TLP250 VCC on each gate driver.
  • 6N137 optocouplers added to every SPI line crossing the isolation barrier (SCK, CS, MISO — one optocoupler per signal per island).
  • MCP3201 sensing rewired: each isolated island has its own MCP3201(s), accessed via the 6N137-isolated SPI.
  • 78L05 5 V regulator on each island to derive the MCP3201 supply from the local 15 V rail.

The intent was to deliver the full isolation architecture that the CHB isolation design note describes.

What failed

Three distinct issues showed up during iteration-3 bring-up. None alone was fatal; together they made the board unreliable enough to drive iteration 4.

1. The 5V_GND ↔ 50V_GND coupling problem (the headline issue)

The iteration-3 layout had inadvertent coupling between the controller's 5V_GND and the bridges' 50V_GND, defeating the otherwise-correct isolation parts. Specifics — reconstructed from the bench notes and the iteration-3 KiCad backup:

  • A continuous inner-plane ground pour extended slightly across the isolation boundary, providing an AC coupling path.
  • The TLP250 and 6N137 footprints had marginal creepage on the PCB-level routing — input-side and output-side copper were close enough to couple parasitic AC current at the cascade-peak common-mode swing.

The symptoms were the classical ones for broken isolation: intermittent false SENSOR_LOST events, garbage STATUS values when Bridge 2 was switching at peak, and occasional protection-trip glitches during clean-load runs.

The fix path — physical separation in the 4-layer stack-up — became the iteration-4 layer plan. See grounding fix for the full story.

2. MISO-topology surprise

The firmware (per Core/Src/spi_mcp3201.c) had been written assuming three independent MISO lines — one per MCP3201. The iteration-3 board surfaced reality: the upper-bridge island has only two MISO data-return lines, with DC2 and current ADCs sharing one (the wire on PC3).

The fix landed in the firmware (the pwm-rewrite-configurable branch) rather than the board:

  • Core/Src/spi_mcp3201.c rewritten to perform a strictly sequential one-channel-per-CS read instead of asserting multiple CS lines and reading three MISOs in parallel.
  • config.h updated: MISO_CUR moved from PC4 to PC3 (shared with MISO_DC2).
  • The build-guide v3.1 section 3.6.2 already described the sequential read sequence — the firmware just needed to match it.

3. The MCP3201 / 78L05 pin-mismatch errata

Ahmet's BUILD GUIDE KICAD MISSMATCH note captures two pin-assignment errors between the build guide and the KiCad symbols used in iteration 3:

Part What was wrong
MCP3201 Build guide had pin 5 = CS, pin 7 = SCK. The actual datasheet has these swapped: pin 5 = SCK, pin 7 = CS. The KiCad footprint followed the datasheet; the build guide had it wrong.
78L05 Build guide had pin 1 = VO, pin 3 = VI. The actual datasheet has these swapped: pin 1 = VO, pin 3 = VI. Same direction of error — schematic followed datasheet, build guide had it wrong.

These were paper errata only — the iteration-3 board was built to the schematic (the datasheet-correct version), and the v3.1 PDF build guide had the misprint. Build Guide v4 corrects both. The same errata page logs the PWM_1L pin error (PA10 in v3.1 → PA12 in firmware and as-built — see pin map).

What was learned

  • Isolation parts don't isolate by themselves. The board's copper layout must respect the isolation boundary just as strictly as the parts do. A continuous inner-plane that "should be ground" can defeat 2.5 kV-rated optocouplers via parasitic AC coupling.
  • Hardware reality wins over firmware assumptions. When the board layout couldn't accommodate three independent MISOs, the firmware was rewritten rather than the board. This is the right call — board respins are expensive, firmware revisions are cheap, and the upstream root-cause (the 3-MISO assumption) was the firmware's responsibility anyway.
  • Schematic-as-source-of-truth pays off. The MCP3201/78L05 pin errors were caught because the schematic followed the datasheet, not the build guide. The build guide is documentation; the schematic is the authoritative source for what gets fabricated.

What changed for iteration 4

Iteration 4 (the as-built) made these specific changes:

  1. Board topology: split the single dual-bridge PCB into two identical single-bridge PCB modules, cascaded externally.
  2. Stack-up: moved from 2-layer to 4-layer with the dedicated, separated ground plane plan (see grounding fix).
  3. MOSFETs: IRFZ44N → IRFB4110, resolving the VDSS headroom + TVS-mismatch issue from iteration 1.
  4. Modulator: kept IPD as the default for a while, then rewrote to PSC in the pwm-rewrite-configurable firmware branch, resolving the bridge-thermal-asymmetry issue from iteration 1.
  5. Dead time: raised from 2 µs (IRFZ44N) to 3 µs (IRFB4110) to accommodate the higher gate charge.
  6. Build Guide v4 published with the errata from this iteration baked in.

Iteration 4 is the as-built. See iteration 4.