MSc Electronics · Edinburgh
30 May 2026
The Idea

You describe the test. The system does the rest.

Run wafer tests by typing plain English. The AI plans the test, the hardware runs it, the data plots itself, and you get a report.

AI-Enhanced Automated On-Wafer Parametric Test System
StudentFrank
SupervisorDr Stewart Smith
LABIIUM leadDr Danial Chitnis
LabSMC Measurement Lab
The shift

From a fragile Windows tool to a system anyone can drive

Today
  • Windows only, with heavy Keysight drivers to install
  • One loose GPIB cable and the whole app crashes
  • Dozens of boxes to fill in by hand for each test
  • Connection errors like VI_ERROR_RSRC_NFOUND on repeat
Where I am taking it
  • Runs on any computer, with no drivers to install
  • The hardware sits behind a small web API
  • You just say what you want to test, in plain words
  • The AI checks it is safe before anything runs
What I am building

Three pillars

The old system already works. I am taking it further in three clear directions.

01

A better interface and charts

Rebuild the front end and make the data come alive: live plots, interactive IV and CV traces, easy result browsing.

02

One hardware layer for everything

Use the LABIIUM bridge instead of low-level drivers, so the same code talks to any instrument, from any computer.

03

AI across the whole flow

An assistant that plans the test, sets the pins, checks safety, runs the maths in code, and writes the report.

The base I start from

A working system that already adapts to any chip

Files mapped
17
Code links
181
Status
Cloned, mapped and running today
Python 3.10 · Streamlit · PyVISA
New chip? Drop in the layout file.
layout.GDS
Industry chip layout
auto-detect pads
Finds test points and their X, Y
probes move
Ready to test the new chip

Pins and test settings live in simple JSON files, so nothing is hard-coded. Point it at a new chip and it adapts. This is the part worth keeping, so I will protect it through the rebuild.

The LABIIUM bridge

No drivers. Any computer. Just a web request.

Your computer
Windows macOS Linux
Sends plain JSON over HTTP
── HTTP / WebSockets ──▶
LAMB bridge
Raspberry Pi 4 box
Holds all the drivers
──── GPIB / USB ────▶
Lab instruments
HP 4142B SMU
HP 4085M switch matrix
+ PA200 probe station
Backed by Keysight Technologies · built by Dr Danial Chitnis at Edinburgh
How a test runs

One loop, start to finish

1You ask

"Sweep the IV curve from 10 V to minus 10 V."

2AI plans

Picks the pins and writes a test task as JSON.

3Sent over the web

A web request goes to the LAMB bridge.

4Instruments measure

Probes touch the wafer and record current and voltage.

5Data comes back

Clean JSON returns over the same web link.

6Charts and report

It plots live and the AI writes up the result.

Build anywhere, then plug in the real lab

Develop on a laptop. Flip one switch for the real instruments.

Top layer · the brain

AI assistant

Understands the request, plans, watches for problems, writes the report.

▼ calls clean APIs ▼
Middle layer · the logic

Test automation

Finds coordinates, maps devices, runs the test steps as plain Python.

▼ simple commands, e.g. set_voltage ▼
Bottom layer · the hardware

LABIIUM bridge

Hides the drivers and talks to the real instruments.

Same code, two modes

Mock or Live

Mock mode. A small simulator makes realistic test data from physics formulas, so almost all of the work (AI flow, layout parsing, curve fitting) is built and tested on a normal laptop, with no instruments.

Diode IV curve SIMULATION = True
0 0.8 V 1.5 V 0 5 10 mA

To use the real lab, flip one setting (SIMULATION = False) and the same code talks to the instruments.

AI smarts

Teaching the AI a lab it has never seen

A model on its own does not know the instruments, is shaky at maths, and cannot read your chip names. These six habits fix that.

A

Knows the physics

Key device formulas (Ohm's law, MOS current, Kelvin resistance) are given to it up front.

B

Never does maths in its head

It must write Python and run it for any number. No guessing, no made-up results.

C

Learns the instruments

Asks which SMU and DMM are connected, looks up their limits, sets a safe current.

D

Learns your chip names

Guesses the naming pattern from the layout, shows you, and fixes it if you correct it.

E

Suggests the right test

A resistor gets a current sweep, a diode gets a voltage sweep. Tweak it by chatting.

F

Remembers, in one file

Everything it learns is saved to a simple local Markdown file. No cloud, easy to track in Git.

Try it · plain words become a test plan pick a prompt
You say

"Test the IV curve of a Kelvin resistor, 0 to 1 mA."

Safety check

OK · 1 mA is within the safe limit

It writes the test task
{
  "device": "R_Kelvin_W10_L100",
  "method": "four_wire_resistance",
  "source": "current",
  "sweep": { "from": 0, "to": 1e-3, "unit": "A" },
  "measure": "voltage",
  "compliance": "10mA"
}
The plan

Four steps to get there

01Now

Clean up the code, build the simulator

Split the logic out of the interface and run a mock lab on a laptop.

02

Plug in LABIIUM, test for real

Use the LAMB bridge to drive the real instruments in the SMC lab.

03

Give the AI tools and guardrails

Let it call the test steps, with hard safety limits and a knowledge base.

04

Full loop and auto reports

From plain words to a finished, plotted report with no manual steps.

To talk about today

Three things I want to ask

Q1

Can we get LABIIUM's web API details from Dr Chitnis (how to send and read data)?

Q2

May we use a LAMB bridge to drive the SMU and switch matrix in the SMC lab?

Q3

Is plain-English control plus auto reports the right focus for the AI work?