Agentic AI for photonics measurement. Fully air-gapped.
Select a stage
…and back to the design intent.
AxMeasureLab Measurement AutomationNo workflow open
Ask the Lab Assistant below to set one up.
Lab Assistant
Ask the lab assistant to control instruments or analyse data.
AxMeasureLab Measurement Automation
The workbench needs a wider screen than this. Open the page on a laptop to watch the OpenTAP run, browse instruments, and talk to the Lab Assistant.
Agentic
AxMeasure plans the test, drives the instruments and extracts the device parameters from the data. The engineer reviews each step. Nothing runs unseen.
Open architecture
Existing instruments, external software and in-house scripts attach to the same workflow and are driven from the same canvas. Probers, tunable lasers, network analysers and source-measure units are a few of the driver categories. Nothing has to be replaced to be included.
Post processing
Raw traces become device parameters in the same run. Resonances are detected, figures such as Q, FSR, group index and loss extracted, and physical models fitted with confidence intervals. Where more than one model could explain a trace, they are ranked on AIC and BIC.
Photonic models
Axiomatic AI's PIC models, and a lab's own models alongside them, sit next to the data, ready for fitting and simulation the moment a measurement finishes.
Local inference
Measurements, layouts and the questions asked about them never leave the building. Agents, instrument control and model fitting run on the on-site control unit, air-gapped. There is no cloud dependency to review and no outbound connection for a security team to approve.
SDK
Every workflow built on the canvas can be generated as Python. A first working script comes out in minutes and stays readable, so changing it later is an edit rather than a rewrite. The SDK gives the same control of instruments and analysis steps, so automation is never blocked by the interface.
At wafer level, the RF test plan has to be written for that DUT's specs: dozens of steps plus the calibration routine, all scripted by the engineer. Enter the specs and AxMeasure returns the plan and the calibration routine in a few minutes.
Efficiency gain
About 5×
A 15-step test plan takes 45 to 75 minutes to script, at 3 to 5 minutes a step. AxMeasure produces it in under 10 minutes.
From the electro-optic test plan, through driving the instruments, to a test conclusion and the plots: AxMeasure runs the whole cycle on the unit in the lab.
Time saved
A full cycle under 10 min
Plan, instruments, conclusion and plots in one local run, in under 10 minutes.
Measured traces are fitted against Axiomatic AI's physical device models in the run that produced them. The model is proposed automatically, but the parameters come from the fit and are scored against the data on AIC and BIC. A result is accepted because it agrees with the measurement, not because a model proposed it.
Steps removed
No export, no second tool
The trace, the derived figures and the fitted model come out of one pass. Nothing is exported to a fitting tool and no parameters are re-entered.
PublicationMay 2026
CLEO 2026, Long Beach · Paper AM2B.6, Metrology for Manufacturing
A Model Context Protocol suite that lets AI agents characterise photonic devices on their own: 39 verified tools drive probers, source-measure units, vector network analysers and optical spectrum analysers, with the results fed back into design.
Safari, Holtorf, Schäfer, Mynampati, Huembeli, Han, Wang, van der Vegt, Radican, Chaudary, Ghadimi, Chen, Taylor, Englund
Conference demoMarch 2026
OFC 2026, Los Angeles · MPI Corporation, booth 502
Unattended measurement on a Lightium thin-film lithium niobate wafer, in front of visitors: test objectives set in natural language, probe coordinates taken straight from the GDS, spectra swept and fitted for Q-factor, group index and propagation loss, with every step traceable to raw data.
PreprintFebruary 2026
arXiv 2602.20144 · with Duke University and MIT
AgentOptics, a framework for autonomous control of heterogeneous optical hardware over the Model Context Protocol. 64 standardised tools across eight devices and a 410-task benchmark, on which agents reach 87.7 to 99.0 percent task success against at most 50 percent for code-generation baselines.
Wang, Han, Cheng, Huang, Ji, Wu, Safari, Holtorf, AlQubaisi, Linke, Zhuo, Chen, Wang, Englund, Chen
Conference demoOctober 2025
ECOC Exhibition 2025, Copenhagen
Optica's CTO filmed a three-way conversation on the exhibition floor with MPI Corporation, Lightium and Axiomatic AI on why wafer-level photonic test has to become automatic, and how AxMeasure configures the probe station, selects devices and orders measurements from a natural-language brief.
Posted by Jose Pozo, Chief Technology Officer, Optica
