Explained · 04 · Conference paper · CLEO 2025

Filming a magnetic field as it changes

Walsh Imaging: a way to recover a magnetic signal’s shape in time and space, without knowing it in advance.

In one lineWalsh sequences let a diamond camera recover arbitrary, unknown magnetic signals in time, space and direction, and a learning algorithm can tune its sensitivity.

The problem

Diamond quantum microscopes can take pictures of magnetic fields with nanotesla sensitivity, which is useful for studying cells, chips and circuits. But most of them rely on constant fields, or on signals that oscillate at one known frequency. A real signal, such as a firing neuron or a switching transistor, is neither. Its shape in time is arbitrary, and you often do not know it beforehand.

The idea: borrow from signal processing

Think of how a piece of music can be written as a sum of pure tones. Walsh functions are a cousin of that idea built from square waves that flip between +1 and −1. Any signal can be described as a sum of them, and, helpfully, a quantum sensor can be made to “listen” for one Walsh function at a time simply by timing the flips (π-pulses) that we apply to its spins.

In Walsh Imaging, we apply a series of these Walsh-patterned pulse sequences to a diamond sensor and record each result with a lock-in camera, which gives a picture across the whole sensor at once. Each run measures one Walsh component of the signal at every pixel. Collect enough components and you can rebuild the full signal over time, and do it for all four NV directions to recover the field’s direction too.

(a) The imaging setup: laser, diamond sensor and lock-in camera. (b) NV center structure and levels. (c) The protocol as a quantum circuit with one qubit per sensing direction. (d) A custom circuit board and a simulation of the microwave field it delivers.
(a) The imaging setup: laser, diamond sensor and lock-in camera. (b) NV center structure and levels. (c) The protocol as a quantum circuit with one qubit per sensing direction. (d) A custom circuit board and a simulation of the microwave field it delivers.

More components, better picture

In simulation, reconstructing an arbitrary waveform with 6 Walsh components captures its rough shape (relative error 0.87), while 50 components follows it closely (relative error 0.27).

Letting an algorithm tune the sensor

Real sensors are imperfect. The microwave field varies across the diamond, and noise limits sensitivity. So we added reinforcement learning, the kind of algorithm that learns by trial and reward. Here an agent chooses the microwave pulse sequences on the independent control lines of the circuit board and is rewarded when the measured sensitivity map approaches a target, so it can adapt the sensor’s sensitivity in space and time while it runs.

(a) The pulse sequence and camera readout for one Walsh component. (b) Walsh functions as a sensing basis. (c) Simulated reconstruction with 6 and 50 components. (d) The reinforcement-learning loop that designs the pulse sequences.
(a) The pulse sequence and camera readout for one Walsh component. (b) Walsh functions as a sensing basis. (c) Simulated reconstruction with 6 and 50 components. (d) The reinforcement-learning loop that designs the pulse sequences.

The limits

This work reports the scheme and simulations. It is a concept paper: the benefit of the learning-based tuning and the full imaging protocol still has to be shown on a real device. The method was later covered by a patent application, described in a separate explainer.