Reading the brain with atoms

Jul 29, 2026 · #neurotechnology #quantum-sensing #magnetoencephalography #neuroscience

Magnetoencephalography spent fifty years requiring liquid helium, a rigid one-size helmet and a subject who does not move. Optically pumped magnetometers put the sensors directly on the scalp of a person who walks, and the gain in signal is real, but every constraint that was lifted reappeared somewhere else.

Essay cover: Reading the brain with atoms, showing optically pumped magnetometers placed on a scalp

A six-year-old is being worked up for epilepsy surgery. The question is where the seizures start, and one of the instruments that can answer it measures the magnetic field produced by the currents inside her cortex. To use it she has to sit inside a helmet that was moulded for an adult head, hold still for an hour, and not fidget by more than a few millimetres. Between her skull and the nearest sensor there are three to four centimetres of nothing: a vacuum gap, a wall of thermal insulation, and a bath of liquid helium at four kelvin. She is six. The recording is often abandoned.

That gap is the whole story. Not the physics of the brain, not the mathematics of source reconstruction, but a few centimetres of empty space that exist for one reason only, which is that the sensor has to be kept very cold and the child does not.

Nine orders of magnitude

The magnetic field escaping a working cortex is somewhere between a hundred femtotesla and a few picotesla. The static field of the Earth, in the same room, is around fifty microtesla. That is a ratio of roughly one hundred million to one. A lift moving two floors below, a car in the street, the steel reinforcement in the wall, all of them produce disturbances thousands of times larger than the thing being measured. Every design decision in magnetoencephalography descends from this single number.

The instrument that made the measurement possible in the first place was the superconducting quantum interference device. A loop of superconductor threaded by a magnetic flux carries a current that depends on that flux in a way that is quantised, and the quantisation gives an exquisitely fine ruler. Noise floors of a few femtotesla per root hertz have been routine for decades. The price is superconductivity, and superconductivity means cryogenics, and cryogenics means a dewar. A dewar has walls. Walls have thickness. Sensors that live inside a shared thermos flask cannot be individually placed, so the helmet is rigid, one size, and generously oversized so that every head fits. Fifty years of functional neuroimaging with millisecond resolution were therefore performed on immobile adults, and the field quietly organised its questions around what an immobile adult can be asked to do.

The atom as a magnetometer

The alternative is to stop cooling anything and let a warm vapour of atoms do the measuring.

Take a glass cell a few millimetres across, put a droplet of rubidium-87 in it, and heat it until the metal evaporates into a dense vapour. Shine circularly polarised laser light through it at 795 nanometres. Because of the way angular momentum is exchanged between photon and atom, the light drives the electron spins of the vapour into one particular state, and once they are all sitting in that state they can no longer absorb photons of that polarisation. The vapour turns transparent. This is optical pumping, and it is a century-old idea being used as an instrument.

Now apply a magnetic field across the cell. The spins are magnetic moments with angular momentum, so they do not simply line up with the field: they precess around it, like a spinning top that leans and then sweeps out a cone rather than falling over. The rate of that precession is the Larmor frequency, and it is proportional to the field:

Read that as a conversion rate rather than an equation. The atom is a tachometer whose reading is the field: leave it in the field of the Earth and it spins at about 350 kilohertz, put it somewhere a thousand times quieter and it slows down proportionally. What the photodiode on the far side of the cell actually sees is the consequence of that precession. Spins that have tilted away from the direction of the beam start absorbing again, so the transmitted light dims. Light in, light out, and the difference is a field measurement.

The trick that makes this competitive with a superconductor is called the spin-exchange relaxation-free regime. Atoms in a dense vapour collide constantly, and each collision scrambles the spin of the pair, which normally destroys the coherence you are trying to read. But if the field is small enough that precession is much slower than the collision rate, the atoms average over their collisions instead of being ruined by them, and the relaxation term nearly vanishes. Sensitivity jumps by orders of magnitude, into the range of tens of femtotesla per root hertz, close enough to a cooled superconducting loop that the difference no longer decides anything.

The catch is written into the same sentence. That regime exists only near zero field. Sensitivity was bought with dynamic range, and we will come back to what that costs.

No cryogenics anywhere in the chain. The cell runs hot, around 150 degrees Celsius, which is the opposite problem from a superconductor and turns out to matter for a different reason.

Why three centimetres matter more than the noise floor

A cooled superconducting loop is marginally quieter than a warm vapour cell. If sensitivity were the only axis, there would be no story here. The reason the vapour wins is geometry.

The field produced by a current dipole in a conducting head falls off steeply outside it, roughly as the inverse square of distance for the overall amplitude. The more important effect is subtler. A field pattern can be decomposed into spatial frequencies, and the high spatial frequencies, the fine structure that distinguishes one gyrus from its neighbour, are attenuated exponentially with distance. Move the sensor away from the scalp and you do not merely receive a fainter copy of the same picture. You receive a blurred one, and the blurring is irreversible. Three centimetres of vacuum is a low-pass filter that no amount of post-processing undoes.

Bring the sensor down to a few millimetres from the skin and both effects reverse at once. Head-to-head comparisons between on-scalp arrays and cryogenic systems report interictal epileptic spikes with amplitudes several times larger, and signal-to-noise ratios higher by tens of percent, with localisation value that is comparable rather than degraded. The gain is largest exactly where the old instrument was worst, which is on small heads: a child inside an adult helmet loses more than an adult does.

The amber dot is a cortical source. On the left it is seen through a low-pass filter made of empty space, on the right it is seen from a few millimetres away.

Where this stands in 2026

The clinical case is being built on epilepsy, because epilepsy provides an unambiguous target: interictal discharges are large, brief and localised. A 2025 study in Epilepsia recorded eleven patients with temporal lobe epilepsy for an hour each, using an on-scalp array with sensors placed on the scalp and additionally on the face to catch basal temporal sources, and reported a mean of thirteen discharges per patient at a mean amplitude of 3.3 picotesla with a signal-to-noise ratio around 9.4. A companion line of work compared the new sensors against superconducting ones for detecting and localising the same discharges and found no significant difference in detection. That negative result is the important one. A new instrument that merely matches the old one, while removing the refrigerator, has already won.

Channel count was the other standing objection, and it is closing. Whole-head systems with eighty and with a hundred and twenty-eight sensors are now in use for cognitive research, which is the density at which source reconstruction stops being an argument.

The change that is qualitative rather than incremental arrived in January 2026, with a platform that integrates a low-noise head-mounted display and six-degree-of-freedom optical motion tracking of head and hands into a wearable magnetoencephalography setup, driven by a game engine. Combined with earlier demonstrations of recordings during walking and during two people interacting, and with battery-powered mobile systems that carry their own lightweight shielding, this moves the experimental vocabulary. For fifty years the price of millisecond resolution was a subject bolted to a chair staring at a fixation cross. It is now possible to ask what a brain does while its owner reaches for something, in a space it can move through.

What breaks

The honest part of the story is that every constraint that was lifted came back wearing different clothes.

The dynamic range of an alkali vapour sensor is around five nanotesla, which is minute. A background field drift of roughly three and a half nanotesla is already enough to cost about five percent of gain error. So freedom of movement is not a property of the sensor: it is a property of the volume in which the residual field stays small. What actually buys movement is a shielded room plus active nulling coils plus, increasingly, motion-adaptive compensation that updates the coil currents in real time. The helium dewar was replaced by an electromagnetic cage that is less visible and no less real.

Bandwidth is the second bill. Alkali cells in the relaxation-free regime work comfortably over roughly one to a hundred hertz. High-frequency oscillations, which are among the more interesting candidate biomarkers in epilepsy, live above that band. Other cell chemistries, helium in particular, trade sensitivity for bandwidth and are being pursued for exactly this reason.

Then there is the crowd problem. Every sensor carries its own coils, and neighbours sitting a centimetre apart perturb one another. Careful coil geometry pushes cross-talk down to a couple of percent, which is a figure that has to be measured and maintained rather than assumed. Triaxial sensors, which reject external interference far better because they see the full field vector, split the laser beam three ways and therefore sit on a higher noise floor. There is no free axis.

Heat is a quiet irony. The vapour cell runs near 150 degrees Celsius a few millimetres from skin. Insulation solves the safety question by adding standoff, which is precisely the quantity the whole design exists to minimise.

The subtlest cost is bookkeeping. Source reconstruction requires knowing where each sensor is and how it is oriented, to something like a millimetre and a degree, relative to the anatomy of the individual. In a rigid cryogenic helmet that information was free and constant. On a flexible cap sitting on a head that moves, co-registration becomes the dominant error term, and a beautifully clean recording can still be localised to the wrong gyrus.

And under all of it sits a ceiling that no engineering removes. What is measured is the summed post-synaptic current of tens of thousands of pyramidal neurons that happen to be aligned with one another, mostly in the walls of sulci. Sources oriented radially to the skull are close to silent. The inverse problem stays formally ill-posed: an infinity of internal current distributions produce the same field outside, so every source map published is a measurement fused with a prior, and the prior is doing work that the picture does not show.

What the removal of a refrigerator actually changes

It would be easy to file this as a cost reduction. It is not. The requirement to hold still was never a fact about the brain, it was a fact about cryogenics, and it silently shaped half a century of questions. Infants and small children were largely excluded. Motor control was studied by asking people not to move. Interaction between two humans was almost unstudiable at this timescale.

At the same time the new instrument draws the boundary of non-invasive reading more sharply than the old one did. It is now realistic to follow a moving brain at millisecond resolution, at the spatial scale of neural populations, through a shielded room and no further. Anything that requires individual cells still requires crossing the skull. The vapour cell has moved the line, and made unusually clear where it now runs.