Skip to content
Contact

Client project · AI · RF measurement

VNA software with an AI assistant: Measure, model, understand.

Software for vector network analysers (VNA). It captures the S-parameters of RF components and stores them in Touchstone format. The measurements become models of the components, managed centrally and under version control, so that every version leads back to its measurement data. An AI assistant answers questions about measurements and models in plain language and points out anything unusual.

Discuss a similar project What a VNA measures

Key facts

Client
Company in radio-frequency engineering
Project
VNA software with an AI assistant and model management for RF components
Technologies
  • VNA
  • S-parameters
  • Touchstone
  • RF models
  • AI assistant

VNA display

Calculated example

Marker 1

Frequency
2.120 GHz
|S21|
−5.92 dB
Return loss
7.39 dB
VSWR
2.49

Model BP-01

Version 3

from measurement M-03

File
m-03.s2p
Checksum
60947e4bd6a7

AI note

Narrow dip at 2.120 GHz. |S21| is 5.10 dB lower there than in the measurement behind version 2.

Based on

  • m-03.s2p
  • m-02.s2p

The row in the file

Calculated example: a Chebyshev band-pass filter with three resonators, centre frequency 2.000 GHz, relative bandwidth 20%, ripple 0.1 dB, unloaded Q 100. Computed, not measured: the measurements, model name, checksum and note all come from formulas. They show how measurement, model and note can fit together, not the client’s components, data or the interface of the software.

The picture shows the calculated example measurement M-03 with 401 points from 1.000 GHz to 3.000 GHz. One chart shows |S21| and |S11| in dB over frequency, a Smith chart shows S11. Marker 1 sits at 2.120 GHz: |S21| −5.92 dB, return loss 7.39 dB, VSWR 2.49. There is a narrow dip at that point, and |S21| is 5.10 dB lower than in the measurement behind version 2. The card names model BP-01 in version 3, made from measurement M-03, and the AI note points to the dip.

The project

From the sweep on the VNA to the answer from the AI. All in one piece of software.

elsyQ built the software for a company in radio-frequency engineering. The result comes down to three points. The sections below explain the terms behind them, with a calculated example instead of client data.

What a VNA measures

One wave in, two answers out: what comes back and what gets through.

A vector network analyser sends a radio-frequency signal into a component and sweeps across a range of frequencies, point by point. At every port it measures the incident and the outgoing wave, in magnitude and phase. That is what vector means. The ratios of these waves are the S-parameters, one set for every frequency point. A filter with an input and an output is a two-port and has four of them.

Four S-parameters, one two-port

Schematic
S11 Reflection at port 1
The ratio of the wave coming back out of port 1 to the wave going in, in magnitude and phase. Port 2 is terminated in a matched load, so nothing runs into port 2.
In the calculated example, measurement M-02 at 2.000 GHz: −33.47 dB
S21 Transmission from port 1 to port 2
How much of the wave going into port 1 comes out at port 2, in magnitude and phase. For a filter, this is the passband curve.
In the calculated example, measurement M-02 at 2.000 GHz: −0.69 dB
S12 Transmission from port 2 to port 1
The same path in the opposite direction. For reciprocal components such as passive filters, S12 equals S21. An amplifier, by contrast, is not reciprocal.
In the calculated example, measurement M-02 at 2.000 GHz: −0.69 dB
S22 Reflection at port 2
The reflection at port 2. This time port 1 is terminated in a matched load, so nothing runs into port 1.
In the calculated example, measurement M-02 at 2.000 GHz: −33.47 dB
Si⁣j=biaj,ak=0(k≠j)

Schematic. The example values come from the calculated band-pass filter on this page, measurement M-02 without the dip, at its centre frequency.

Schematic of a two-port with port 1 and port 2. A wave runs into each port, a1 and a2, and one runs out of each, b1 and b2. S11 is b1 divided by a1 and S21 is b2 divided by a1, both measured while nothing runs into port 2. S12 is b1 divided by a2 and S22 is b2 divided by a2, measured while nothing runs into port 1.

The formulas behind it

Z0 is the reference impedance, in RF work usually 50 Ω. Return loss and VSWR depend only on the magnitude of Γ; the phase is carried by Γ alone.

S-matrix
(b1b2)=(S1⁣1S1⁣2S2⁣1S2⁣2)⁢(a1a2)

It links the incident waves a to the outgoing waves b, afresh for every frequency point.

Reflection coefficient
Γ=Z−Z0Z+Z0

How strongly a load Z reflects a wave, relative to Z0. If Z equals Z0, Γ is zero and nothing comes back. S11 is the reflection coefficient at port 1.

Return loss
RL=−20⁢log10⁡|Γ|

The reflection in dB. The larger the value, the less comes back: at 20 dB, one per cent of the power is reflected.

Voltage standing wave ratio
VSWR=1+|Γ|1−|Γ|

The ratio of the highest to the lowest voltage of the standing wave that forms on the line when the load is mismatched. A value of 1 means a perfect match, infinity means total reflection.

Insertion loss
IL=−20⁢log10⁡|S2⁣1|

How much weaker a wave arrives at port 2 than it was when it went into port 1, in dB. For a filter it is small in the passband and large outside it.

The Smith chart, hands on

Every load has its point. At the centre, nothing comes back.

The Smith chart shows the reflection coefficient Γ as a point inside a circle. At the centre Γ is zero; on the rim everything is reflected. The short circuit sits on the left, the open circuit on the right, inductive loads in the upper half and capacitive ones in the lower half. The circles mark constant resistance and the arcs constant reactance, so the impedance can be read straight off the chart. The distance from the centre shows how badly a load is matched, and the angle gives the phase of the reflection.

One load, one point

Calculated exactly
Load Z
50 + j50 Ω
Normalised z = Z/Z0
1.00 + j1.00
Reflection coefficient Γ
0.200 + j0.400
Magnitude |Γ|
0.447
Angle of Γ
63.4°
Return loss
6.99 dB
VSWR
2.62

The seven example loads, calculated exactly with Z0 = 50 Ω
Load Impedance Γ |Γ| Angle Return loss VSWR
Matched 50 Ω 0.000 + j0.000 0.000 undefined ∞ 1.00
Open ∞ 1.000 + j0.000 1.000 0.0° 0.00 dB ∞
Short 0 Ω −1.000 + j0.000 1.000 180.0° 0.00 dB ∞
Below Z0 25 Ω −0.333 + j0.000 0.333 180.0° 9.54 dB 2.00
Above Z0 100 Ω 0.333 + j0.000 0.333 0.0° 9.54 dB 2.00
Inductive 50 + j50 Ω 0.200 + j0.400 0.447 63.4° 6.99 dB 2.62
Capacitive 50 − j50 Ω 0.200 − j0.400 0.447 −63.4° 6.99 dB 2.62

Same distance, opposite side: 25 Ω and 100 Ω are equally badly matched, both with a VSWR of 2.00 and a return loss of 9.54 dB. Only the angle of Γ tells them apart. A scalar analyser, which measures magnitude only, sees no difference here; a vector analyser does.

A Smith chart with the point of the selected load. The default is 50 + j50 Ω: Γ lies at 0.200 + j0.400, magnitude 0.447, angle 63.4°. The table lists the values of all seven example loads.

Touchstone, decoded

One measurement, one text file. Readable by people and by programs.

The software stores the measurements in Touchstone format. Touchstone is a widely used text format for S-parameters that people can read in an editor and simulators can import directly. The file extension gives the number of ports: .s1p for a one-port, .s2p for a two-port. The line starting with # is the option line, and everything after a ! is a comment. For one-ports and two-ports, one line then follows for each frequency point, holding all the S-parameters.

Excerpt from a measurement file

Calculated example
File
m-03.s2p
Checksum (SHA-256)
60947e4bd6a7

401 data lines

# GHz S DB R 50
1.000    -0.015   165.05   -48.112  -106.11   -48.112  -106.11    -0.015   165.05
1.005    -0.016   164.92   -47.893  -106.25   -47.893  -106.25    -0.016   164.92
! ... 220 lines omitted
2.110   -12.178   -30.93    -2.931   -66.96    -2.931   -66.96   -12.178   -30.93
2.115    -9.166   -47.51    -4.573   -67.44    -4.573   -67.44    -9.166   -47.51
2.120    -7.391   -69.05    -5.921   -56.33    -5.921   -56.33    -7.391   -69.05
2.125    -7.074   -89.95    -4.980   -44.20    -4.980   -44.20    -7.074   -89.95
2.130    -7.663  -106.00    -3.514   -42.73    -3.514   -42.73    -7.663  -106.00
! ... 173 lines omitted
3.000    -0.056  -152.45   -32.553   119.68   -32.553   119.68    -0.056  -152.45

The option line

#
Starts the option line
GHz
Unit of frequency
S
Kind of parameter, here S-parameters
DB
Each number pair: magnitude in dB, angle in degrees
R 50
Reference resistance in ohms

If an entry is missing, the default applies: GHz, S, MA and R 50.

The columns

Frequency
Frequency of the point, in the unit given by the option line.
S11
Reflection at port 1, as a number pair in two columns.
S21
Transmission from port 1 to port 2, as a number pair.
S12
Transmission from port 2 to port 1, as a number pair.
S22
Reflection at port 2, as a number pair.
  • A quirk of the format: for two-ports, S21 comes before S12. From three ports upwards, the values of the matrix follow row by row.
  • The decimal separator in the file is always a point, whatever the language.
  • In the example, S12 equals S21 and S22 equals S11: the calculated filter is reciprocal and symmetrical.
  • The highlighted row is the one for 2.120 GHz. That is where the dip lies that the AI note in the example points to.
  • The lines starting with ! ... are cuts made for this page. As comments, they do not break the format.

Calculated example: 401 data lines, one per frequency point from 1.000 GHz to 3.000 GHz. Computed, not measured, and not a file from the project. The checksum is the SHA-256 hash of the whole file in the dB-and-angle format, shortened here to twelve characters.

Models with provenance

No model without a measurement. Every version leads back to its data.

The measurements become models of the RF components. The software manages them centrally and under version control, and every version leads back to its measurement data. The illustration below shows in general terms why that matters: comparing two versions means comparing the measurements behind them, and seeing where and by how much they differ.

Model BP-01 in three versions

Illustration · calculated data

  1. Version 1

    First measurement

    from measurement M-01

    201 points · 1.500 GHz to 2.500 GHz

    File
    m-01.s2p
    Checksum
    83e9811d4997

  2. Version 2

    Measurement over a wider frequency range

    from measurement M-02

    401 points · 1.000 GHz to 3.000 GHz

    File
    m-02.s2p
    Checksum
    7721f71af0bd

  3. Version 3

    current

    New measurement with a narrow dip

    from measurement M-03

    401 points · 1.000 GHz to 3.000 GHz

    File
    m-03.s2p
    Checksum
    60947e4bd6a7

Compare two versions

The measurements behind versions 2 and 3 differ most at 2.120 GHz, by 5.10 dB. Between 2.105 GHz and 2.140 GHz the deviation is above 1 dB, everywhere else it stays at 0.87 dB or less.

Illustration with calculated data: the model name, measurements, files and checksums are examples. The curves are the measurements behind the versions. They show what versioning with provenance achieves, not how the software builds or stores the models.

BP-01, three versions, each with its measurement file and checksum. The default comparison shows |S21| of the measurements behind the two latest versions from 1.700 GHz to 2.300 GHz: the measurement behind the latest version lies 5.10 dB lower at 2.120 GHz.

Traceable

Every version comes with the measurement file it was made from. A checksum shows whether the file is still the same.

Comparable

Two versions can be compared point by point, because both go back to measurement data on a frequency axis.

Nothing is overwritten

With versioning, a new version does not replace the old one; it is added. The old one stays, together with its data.

The three points describe versioning with provenance in general. How the software builds, stores and compares the models internally is not described on this page.

AI assistant

Questions in plain language, and a pointer to anything unusual.

In the software, an AI assistant answers questions about measurements and models in plain language and points out anything unusual. The example below shows what such an exchange can look like, worked out with the example data on this page.

A question for the assistant

Example · calculated data
  • computed deterministically
  • worded by the language model

  1. Question

    Is anything unusual about measurement M-03?

    AI assistant

    Yes. Between 2.105 GHz and 2.140 GHz, |S21| has a narrow dip, deepest at 2.120 GHz with −5.92 dB. In the measurement behind version 2, the value at the same point was −0.82 dB, so the loss there has grown by 5.10 dB. The data alone do not show whether the dip comes from the component or from the measurement set-up. Someone should check that.

    Based on

    How the assistant can go about it

    1. Reads the measurement files of both versions. (computed deterministically)
    2. Compares |S21| point by point, threshold 1 dB. (computed deterministically)
    3. Words the answer and names the spot. (worded by the language model)

  2. Question

    How well is the component matched in the passband?

    AI assistant

    In the passband of the example, from 1.810 GHz to 2.210 GHz, the return loss in the measurement behind version 2 is at least 16.83 dB everywhere. It reaches its lowest value at 2.210 GHz, at the upper edge of the band. In the measurement behind version 3 it drops to 7.07 dB at 2.125 GHz, right where the dip is. The lower the return loss, the more power comes back.

    Based on

    How the assistant can go about it

    1. Reads the measurement data of both versions. (computed deterministically)
    2. Computes the return loss at every point from 1.810 GHz to 2.210 GHz. (computed deterministically)
    3. Words the answer. (worded by the language model)

  3. Question

    What changed from version 2 to version 3?

    AI assistant

    Version 3 comes from measurement M-03, version 2 from M-02. Both measurements cover 1.000 GHz to 3.000 GHz with 401 points. Between 2.105 GHz and 2.140 GHz they differ by more than 1 dB, most at 2.120 GHz with 5.10 dB. Everywhere else the deviation stays at 0.87 dB or less.

    Based on

    How the assistant can go about it

    1. Reads the provenance of both versions. (computed deterministically)
    2. Compares the measurement data point by point. (computed deterministically)
    3. Sums up the difference. (worded by the language model)

Example: the questions, answers and steps are made up and work with the example data on this page. They show how an assistant can work with measurement data, not how the assistant in the project is built. The file names lead to the data on this page.

Numbers from the data

Values such as the loss at a given frequency can be read and calculated exactly from the measurement file. A language model can get that arithmetic wrong. So it makes sense to split the work: code does the computing, and the model only does the wording.

Claims with evidence

An answer that names the file and the frequency range can be checked without having to take the assistant’s word for it.

A pointer, not a verdict

Whether an anomaly comes from the component or from the measurement set-up is for a person to decide. An assistant can show where it is worth looking.

The three points describe how an assistant can work with measurement data in general. What the assistant in the project is built on and how it works is not described on this page.

How elsyQ separates algorithm and language model: Determinism and AI

Technology

Five terms from the project, each explained briefly. Where the key facts say something about their use in the project, it is noted alongside.

VNA

A vector network analyser measures, across a range of frequencies, how a component reflects and transmits radio-frequency signals, in magnitude and phase.

In this project The software brings measurement and evaluation on the VNA into one interface.

S-parameters

Scattering parameters: the ratio of outgoing to incident wave at the ports of a component, as a complex number for every frequency point.

In this project The software captures the S-parameters of the RF components.

Touchstone

A text format for S-parameters, with an option line and the measured values for each frequency point. The file extension gives the number of ports, and a file for a two-port ends in .s2p.

In this project The software stores the measurements in Touchstone format.

RF models

A model describes how a component behaves at high frequencies, in a form that further work can build on, for example a circuit simulation.

In this project The measurements become models of the components, managed centrally and under version control, every version with its measurement data.

AI assistant

An assistant, usually built on a language model, that answers questions in plain language and draws on data to do so.

In this project Answers questions about measurements and models in plain language and points out anything unusual.

RF and measurement at elsyQ

Independently of this project: RF electronics is a focus of our founder, and measurement technology is part of what our team covers. We read schematics like source code and measure before we write firmware for a board. On top of that comes software that computes: numerics, simulation and machine learning on measurement data.

Meet the team See all projects

Frequently asked questions

What is the VNA software with an AI assistant?

The VNA software with an AI assistant is a client project by elsyQ for a company in radio-frequency engineering: software for vector network analysers (VNA). It captures the S-parameters of RF components and stores them in Touchstone format. The measurements become models of the components, managed centrally and under version control, so that every version leads back to its measurement data. An AI assistant answers questions about measurements and models in plain language and points out anything unusual.

What does a vector network analyser measure?

A vector network analyser, or VNA, measures how a component reflects and transmits radio-frequency signals, in magnitude and phase; elsyQ has built software for such instruments in a client project. To do so, it records the incident and the outgoing wave at every port across a whole range of frequencies. The ratios of these waves are the S-parameters.

What are S-parameters?

S-parameters, or scattering parameters, describe how an RF component reflects and transmits waves at its ports; they are exactly what the VNA software that elsyQ built for a company in radio-frequency engineering captures. Each S-parameter is the ratio of an outgoing to an incident wave, as a complex number for every frequency point. A two-port such as a filter has four of them: S11 and S22 for reflection, S21 and S12 for transmission.

What is the Touchstone format?

Touchstone is a widely used text format for S-parameters, and it is the format in which the VNA software that elsyQ built for a company in radio-frequency engineering stores the measurements. An option line names the frequency unit, the kind of parameter, the number format and the reference resistance, followed by the values for each frequency point. The file extension gives the number of ports, and a file for a two-port ends in .s2p.

Why are the models kept under version control?

In the software that elsyQ built for a company in radio-frequency engineering, the models of the RF components are managed centrally and under version control, and every version leads back to its measurement data. In general, versioning with provenance makes results traceable: whoever uses a version can see which measurement it came from and can compare two versions by their data. A new version does not replace an old one; it is added alongside it.

What does the AI assistant do?

The AI assistant in the VNA software that elsyQ built answers questions about measurements and models in plain language and points out anything unusual. The conversation on this page is an example with calculated data and does not show the assistant from the project. How elsyQ generally separates what an algorithm computes from what a language model words is described on the page Determinism and AI.

Does elsyQ also build software for other measurement instruments?

elsyQ builds software for measurement technology beyond this project as well, from driving the instruments to evaluating the data. RF electronics is a focus of the founder, measurement technology is one of the team’s competences, and machine learning on measurement data is one of its services. Which instruments, formats and evaluations you need is something we clarify at the start of the project.

Trademark notice Touchstone is a mark of its owner and is used here only to identify the file format for S-parameters. elsyQ is not affiliated with that owner.

Work with us

Measurements, models and answers under one roof?

Tell us what you measure and what should come out of it in the end. We will tell you how we would build it.

Discuss a similar project All projects