SPICE Simulation Skills

Circuit simulation for AI agents in the Adom ecosystem. Install this pack and your agent knows which SPICE engine to use, where to run it, how to drive it without a person clicking, and what each tool's licence allows. It also installs Adom SPICE, a dashboard that runs a circuit in ngspice, LTspice or PSpice for TI (one, two or all three) and shows the results on one set of axes.

Everything here comes from a hands-on comparison on a real board: a TPS54202 12 V to 5 V, 1 A buck converter molecule.

A 4 minute tour: the board in 3D, the same test in ngspice, LTspice and PSpice for TI, all four runs on one set of axes, the Adom SPICE dashboard, and an AI thread in Adom Hydrogen running the whole comparison from one prompt. Download the video.

Ask your AI

You do not need to learn any commands. This is the prompt typed into a Claude thread in Adom Hydrogen in the video:

Install adom/adom-spice-skillpack from the Adom wiki. Then use the adom-spice dashboard to simulate
the TPS54202 12 V to 5 V buck example in ngspice, LTspice and PSpice for TI (LTspice and PSpice for TI
are installed on <your Windows PC>, so set the bridge target to <your Windows PC>). Show the dashboard
in a webview here and tell me whether the three engines agree.

The agent installed the pack, ran ngspice in its container, drove LTspice and PSpice for TI on the Windows PC through Adom Bridge, opened the dashboard in a Hydrogen webview, and answered with the comparison:

Most of the time, one engine is enough

Running all three engines is how you check a model. Day to day, one engine usually does the job, and ngspice is the default because it runs right in your container:

• "Simulate my buck converter in ngspice with the Adom SPICE dashboard and show me the start-up and load-step response."
• "This regulator's model is encrypted for PSpice. Run it in PSpice for TI on my Windows PC and show me the output ripple."
• "Run my circuit in LTspice as well, and tell me if it agrees with ngspice."

A single-engine run gets the same charts and measurements; add a second engine whenever you want a cross-check.

The Adom SPICE dashboard

adom-spice runs one study in any of the three engines. ngspice runs in your container. LTspice and PSpice for TI run on your Windows PC through Adom Bridge (PSpice for TI has no batch mode, so the dashboard opens it, presses Run, and reads the results back). Every engine is measured by the same code, so when two engines disagree, the dashboard names the measurement and the engine that is out of line. That is usually a netlist-translation or model problem worth finding before you build the board.

It also keeps a ledger of every run: the simulation time (LTspice and PSpice times stay hidden by default, per their licences) and the AI tokens the run cost.

On the example board all three engines agree on all eight measurements: output voltage, output ripple, inductor ripple, start-up peak current, time to 95 %, overshoot, undershoot and settling. TI's own encrypted model, added as a reference, starts up the same way, sits about 30 mV higher, and swings a little more on each load step.

Driving it yourself:

adom-spice --ai-thread "buck check" start
adom-spice --ai-thread "buck check" study ~/.adom/adom-spice/examples/tps54202-buck/study.json
adom-spice --ai-thread "buck check" run ngspice                  # or: run ngspice ltspice pspice
adom-spice --ai-thread "buck check" settings bridgeTarget=<your Windows PC>   # for LTspice and PSpice for TI
adom-spice --ai-thread "buck check" show                         # opens the dashboard in a Hydrogen webview

A study is one JSON file: a netlist per engine, the signal names, the chart windows and the measurements. The adom-spice skill in this pack documents the format.

Which engine to use

• Default to ngspice in the Ubuntu container. It is free (BSD), small, runs headless and in parallel, and works the same in WSL2, in Lima on a Mac, and in the cloud. It is the only mainstream option that may back a hosted simulation dashboard.
• Use PSpice for TI for PSpice-encrypted vendor models (TI and many power-device vendors ship these). Nothing else can run them. It runs on a Windows PC through Adom Bridge, for internal checks.
• Use LTspice for Analog Devices / Linear Technology parts on a Windows PC through Adom Bridge.
• Never decrypt a vendor model. Validate a fast hand-built model against the vendor model once instead.

What we learned

• Accuracy: on the same well-posed netlist the engines agreed closely. The differences we saw came from translating a netlist between dialects, not from the solvers.
• Speed: behavioral models run in seconds. A transistor-level vendor model of a switching regulator takes minutes to tens of minutes for a few milliseconds of switching, so iterate with a behavioral model and validate once with the vendor model.
• Automation: ngspice is the easiest to drive from an agent. LTspice has a working batch mode once its first-run dialogs are answered. PSpice for TI has no batch runner in this edition, so an agent has to drive its window.
• Install friction: ngspice installs in minutes with no account. LTspice installs per-user without admin rights. PSpice for TI needs a TI account, an export-control request, an access key and an admin install a person must approve.
• Model validation: our hand-built model matched TI's vendor model on start-up and ripple but under-predicted load-step deviation by about 15 %. Fitted models are good for start-up and ripple; check their loop dynamics against the vendor model.
• AI cost: a fresh agent running one board simulation used about 40,000 tokens with ngspice or LTspice and about 49,000 with PSpice for TI (it has to drive a window). Most of that is the agent's fixed start-up context, not the simulation.

We do not publish timing or head-to-head numbers for LTspice or PSpice because their licences restrict publishing benchmarks. The ngspice numbers below are our own.

ngspice in the container (measured)

Skills in this pack

Examples

• app/examples/tps54202-buck/: the dashboard study for the board in the video, with one netlist per engine.
• examples/buck_portable.cir and examples/buck_portable_pspice.cir: a simple buck test bench in the ngspice/LTspice and PSpice dialects.
• examples/sallen_key_ac.cir: the Sallen-Key filter used for the Monte Carlo run.
• examples/csdf.py: a reader for PSpice .PROBE/CSDF text output.

Install

adom-wiki pkg install adom/adom-spice-skillpack

This installs the main skill, eight sub-skills and the adom-spice dashboard command. LTspice and PSpice for TI themselves are installed on your Windows PC; the spice-ltspice and spice-pspice-ti skills walk your agent through that.