Coming back for more?
Yes, sucker for punishment. After that complete fiasco trying to run Adom's Hydrogen AI tool with Fusion Electronics, I'll have another go at reproducing John Lauer's demo board - but this time in Kicad.
I've already got the Hydrogen GUI / Bridge installed on my machine, along with the integrated Claude Code / Visual Studio Code function. So it's relatively simple to get it to do an automated, hands free installation of Kicad on my machine and set it up to design a circuit and PCBA.
Again, in my case I want it to use the JLCPCB / LCSC database of compts, so that it only selects compts that are available through my preferred CEM that would be tasked with manufacturing and assembling the PCBA (JLCPCB in China). For now though, I'll stop before the ordering stage (it's only a proof of function test) but I want to see all the different stages being completed successfully.
My starting prompts were:
- set up the KiCad and JLCPCB route on my PC
- replicate the example used by John Lauer - a 12V / 5v buck converter
That's all I gave it, although clearly that original demo came with its own set of parameters and conditions. I'm not bothered to know what those were at this stage - I want to see if it actually completes the process and if so, does it resemble his example and does it look fit for purpose. Then I may get it to develop a schematic and PCB layout for my TIG button circuit, possibly even taking it through production of a small number of boards at LCSC for actual use here.
Terminology:
For reference:
The design tool is Adom AI Flow. Adom and Hydrogen are the levels around it:
- Adom is the company and the overall platform. Everything carries its name: the wiki (wiki.adom.inc), Adom Bridge, Adom Fields.
- Hydrogen, or "Adom Hydrogen", is the desktop app on your PC: the window with the chat on the left and dashboards such as the AI Flow timeline on the right. It's the workspace, not the design tool.
- Adom AI Flow (the adom-aiflow command) is what ran the design. It takes the board from the prompt through parts, schematic, simulation, placement, routing, pours, current and heat checks, silkscreen and fab files, and it produced the timeline and the run video.
- KiCad (or Fusion, in John's second build) is the actual circuit-board editor AI Flow drives, through Adom Bridge. Adom Fields did the current and heat solve.
So for this board you'd say "it was designed with Adom AI Flow in KiCad, run from Hydrogen". There.
And off it went....
No intervention required from me, no crashes, all my monitors etc remained fully functional throughout, other apps continued to work OK etc while it worked in the background.
You can see the steps in the design process here, part way through. Again, I've not looked at what these steps entail in any detail but they appear to be reasonably convincing:
Having the Kicad window open allowed me to see the process evolving - schematic, layout, simulation, verification etc.
I kicked it off at 2:35pm here and it was all done and dusted by 4:54pm ie 2 hours and 20 mins.
It created an AI-generated video of the process that I've uploaded to my channel here:
Here's the layout in 2D with all layers visible:
Top side copper:
Bottom side copper:
The silk screen ident is a bit of a mess, with text on top of pads, so it's not clear at this stage how it passed the JLCPCB DFM rule check. I was actually going to take this further, I'd kick somebody's ass and get them to fix that. But more on this shortly....
And then to the really horny bit - the 3D view.
And flipped over to see the bottom side:
...at which point you realise that half the silk screen characters are on the wrong side of the board, hence there's no text actually placed on the copper pads. Not a lot of use for fault finding or manual placement but at least you can now understand why it passed the DFM check.
How does it compare with John's original design?
Here are John's (top) and mine (bottom):
Beyond the appearances, we have some other differences - and a lot of similarities.
Here's Claude's comparison of the two designs. Note - the table headings are shifted to the left for some reason only known to Blogger:
| John, Fusion | Yours, KiCad + JLCPCB | |
|---|---|---|
| Fab target | His "3rd party fab" rules (0.5 mm-class text) | JLCPCB 2-layer rules (0.8 mm minimum text) |
| Parts | Original parts, including the Abracon AMPLH5030S inductor | JLCPCB/LCSC parts, including the YOUCHI HCMA-0530 inductor; 9 swaps checked against both datasheets |
| Placement | His KiCad placement | The same placement, with J7, TP7, TP10 and R7 nudged to pass the newer checks |
| Heatsink vias at U1 | 6 (after one rework) | 3, plus 2 more at its GND pin; this routing's switch-node trace takes the rest of that pocket |
| Peak current density | 53 A/mm² | 40 A/mm² |
| Hottest rise, still air | U1 +49 °C; page summary says 57 °C against a 55 °C budget | U1 +10 °C and L1 +46 °C by AI Flow's estimate (Fields solve peak +46 °C), all under 55 °C |
| Silkscreen | Framed tags on every contact and corner pin, values beside every reference, identity on the bottom | No frames; nets and references at 0.8 mm, most values left off, small parts' references on the bottom |
| Time, prompt to finish | 2 h 42 min | About 1 h 51 min of AI time |
His heat figure and mine aren't strictly comparable. His Fusion solve also counted heat through the vias and uses Fusion's copper, and the two boards use different inductors.
The most visible difference is the silkscreen. His Fusion board has framed contact tags and a value on every part; yours is plainer because JLCPCB's 0.8 mm minimum doesn't leave room for that on a 28 × 20 mm board.
Conclusion:
It's a bit premature to give out gold stars but at first sight the result looks almost plausible. The tracking looks OK from a distance.
As a proof of function, this is a lot more convincing that yesterday's carry on. I've not used Kicad before and perhaps I don't need to learn how to use it. Although I've done a lot of layouts with Mentor, Altium and EasyEDA (JLCPCB's free tool) in the past, each CAD system takes some learning and Kicad is no toy.
So I may need to try it out for real on my TIG button. Previously, I'd used Claude to design the circuit and do Spice simulations and I was expecting to use the autorouting feature within EasyEDA rather than manually route the board myself. Autorouters tend to be best used for digital buses etc rather than analogue and power electronics circuits but this is a simple, small circuit so what could possibly go wrong?
No comments:
Post a Comment