PYTHON · NUMPY

Learned matching: when AI designs the interface

Hand the matching problem to a network and watch it become a coordinate system a model works in.

01Challenge

Clone an exhaustive search into a small net, and match loads it has never seen in one forward pass.

02Model

The chart began as paper. Then it became software, then it moved onto the die — at millimetre wavelengths on-chip wiring is transmission line, so matching networks are now fabricated in CMOS and tuned while they run. Plasma etch chambers are driven through matchboxes that re-tune themselves mid-pulse as the plasma's impedance moves. Superconducting qubits are read out through impedance-matched amplifiers. Every transistor in the machine you are reading this on was etched behind a live matching loop. The newest turn is that the chart has become a representation to learn in: networks that operate in chart coordinates, because the geometry is a good inductive bias. This module does the smallest honest version of that.

03Guided practice

1. The expert. An exhaustive search over element values that finds the best match for a load. It is correct and slow. 2. The lesson hiding in the labels. Two element values solve each load; if you let the search pick either, the labels flip between neighbouring loads and the net learns their average, which matches nothing. Pin one branch and the map becomes learnable. 3. The clone. A small net maps the load straight to the network. One forward pass, no search, on loads it never saw.

04Feedback

PASS WHEN PASS when the learned net lands near the exhaustive search on held-out loads.

If your run fails, check:

A cell errored — run it top to bottom.
The net should get close to the expert and far ahead of doing nothing.

05Retrieve & space

Why did pinning one solution branch matter? (two valid answers per load make the labels multi-modal; the net learns their average).
Where does impedance matching appear in chip fabrication? (RF matchboxes driving plasma etch and deposition chambers).
Why would a model prefer chart coordinates? (the geometry carries the structure of the problem — a good inductive bias).

06Mastery & project

You can explain why the chart is a useful representation for a learned model, and what makes a matching task learnable.

The habit: when cloning an expert, first ask whether its answer is unique.

GO FURTHER · THE LIVE PLAYGROUND

Smith Chart Navigator — Graph view ↗ — Search 'plasma', 'chiplet' or 'qubit' to trace the chart down the compute stack.

← The physical price of a bit: Landauer Impedance control and contact: staying inside the disk →

PYTHON · NUMPY

Learned matching: when AI designs the interface

Train a net to design matching networks and race it against the exhaustive search it learned from.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you