Brain-Computer Interfaces Can Read Your Neurons. This Printed Circuit Writes to Them.
Northwestern University researchers printed artificial neurons from molybdenum disulfide nanosheets onto flexible plastic film using an aerosol jet printer. The circuits replicate 6 types of biological spiking complexity, operate at frequencies up to 20 kHz, survive more than a million cycles, and successfully stimulated real Purkinje neurons in mouse cerebellar brain slices. It is the first fully printed electronic device to write signals into living brain tissue.
200,000. That is how many synaptic inputs a single Purkinje neuron integrates. Purkinje cells are the sole output neurons of the cerebellar cortex, among the largest and most complex cells in the human nervous system, firing at 50 to 150 Hz in baseline conditions and producing both simple and complex spike patterns that govern motor coordination, balance, and learned movement. On April 15, 2026, a team at Northwestern University demonstrated that a circuit printed on flexible plastic film could make one fire, producing a response in the Purkinje cell that looked indistinguishable from a genuine presynaptic input arriving along a parallel fiber.
Published in Nature Nanotechnology, the paper describes aerosol-jet-printed graphene/MoS2/graphene memristive devices on polyimide substrates. Led by Mark Hersam, chair of materials science and engineering at Northwestern's McCormick School, the team built neuristor circuits that exhibit volatile threshold switching with snap-back negative differential resistance. In plain language: they print layered nanosheets that spike and reset like neurons. Not approximately. Not metaphorically. The waveforms match physiological timescales, and the team demonstrated six distinct types of biological spiking complexity: integrate-and-fire behavior, spike latency, tonic firing, Class 1 excitability, tonic bursting, and phasic dynamics. Each of these corresponds to a recognized classification in computational neuroscience, the Izhikevich taxonomy that catalogues how real neurons encode information.
From Microphone to Speaker
Every brain-computer interface deployed in a human patient reads signals. Every single one. Neuralink's N1 implant, Synchron's Stentrode, BrainGate's microelectrode arrays: all are microphones, translating neural activity into digital commands. Noland Arbaugh can play chess with his Neuralink because the device listens to his motor cortex and interprets intention. What no BCI has done is write back to the brain using electronics that spike with the same waveform vocabulary as the neurons themselves.
Conventional stimulation already exists in clinical use: deep brain stimulation for Parkinson's disease delivers electrical pulses through implanted electrodes, and cochlear implants encode sound as current patterns on an electrode array, but these devices blast coarse electrical signals into tissue, more bullhorn than speaker. A DBS electrode fires at a fixed frequency, typically 130 Hz, with no attempt to replicate the integrate-and-fire dynamics, the bursting patterns, or the phasic responses that neurons actually use to communicate with each other across synaptic gaps. Hersam's printed neurons speak the native language, producing waveforms that a Purkinje cell recognizes as legitimate input.
"We are trying to mimic the brain as faithfully as possible," Hersam told Live Science. "What motivates us is to come up with an alternative to conventional digital computing to handle large amounts of data in a more energy-efficient way."
A Number Nobody Tracks: Energy Per Biologically Compatible Spike
Here is the original calculation that nobody in BCI research seems to be running. A biological neuron fires at approximately 20 femtojoules per spike (10-15 joules), a number so small it barely registers on any instrument humans have built. A single NVIDIA H100 GPU inference token costs roughly 0.002 joules, or 2 × 10-3 joules, which means GPU computation burns one hundred billion times more energy per computational unit than biology does. So where do printed MoS2 neurons fall on this spectrum, and does the answer actually matter for the future of brain-computer interfaces?
Based on the voltage thresholds and currents reported in the Nature Nanotechnology paper, these devices operate in the picojoule-to-nanojoule range per spike (10-12 to 10-9 joules), which places them in a fascinating middle ground: three to six orders of magnitude more energy-efficient than GPUs per computational spike, yet still three to six orders of magnitude hungrier than the biological neurons they mimic. But they can be printed on polymer film with an aerosol jet, mass-produced like circuit boards, and directly interfaced with biological tissue.
| Platform | Energy per spike/operation | Can interface with neurons? | Manufacturable at scale? |
|---|---|---|---|
| Biological neuron | ~20 femtojoules | Yes (is one) | No |
| Printed MoS2 neuron | ~pJ to nJ range | Yes (demonstrated in vitro) | Yes (aerosol jet printing) |
| Intel Loihi 2 (neuromorphic) | ~23 pJ per synaptic op | No | Yes (fab-produced) |
| NVIDIA H100 (GPU) | ~2 mJ per token | No | Yes (fab-produced) |
No published metric tracks "energy per biologically compatible spike" as a benchmark, which is precisely the gap this work exposes. Neuromorphic chips like Intel's Loihi 2 optimize for energy-efficient AI computation but cannot interface with living tissue because they are fabricated on rigid silicon substrates with no biocompatible output stage. BCIs like Neuralink optimize for signal readout but do not generate neuron-native waveforms for stimulation, relying instead on simple current injection. Northwestern's printed neurons occupy a previously empty cell in this matrix: energy-efficient, biologically compatible, and printable at scale using equipment that costs orders of magnitude less than a semiconductor fab.
Why Purkinje Cells Are the Hardest Test
Choosing Purkinje neurons for the demonstration was not arbitrary, and the reasoning reveals something important about what this technology might eventually do. With their 200,000 synaptic inputs, dendritic trees spanning hundreds of micrometers, and dual spiking modes (simple spikes at 50-150 Hz, complex spikes driven by climbing fiber inputs), Purkinje cells are the most computationally demanding neurons in the mammalian brain. They are the sole output pathway of the cerebellar cortex, meaning every motor correction your cerebellum computes exits through Purkinje cell axons. If a printed electronic neuron can drive a Purkinje cell, it can in principle drive simpler neurons elsewhere in the nervous system.
For motor rehabilitation, this distinction between reading and writing matters more than anywhere else in neuroscience. Cerebellar damage from stroke or traumatic brain injury disrupts motor coordination, and no existing BCI addresses cerebellar circuits because the computational complexity of Purkinje cells exceeds what simple electrical pulses can meaningfully engage. A device that speaks the Purkinje cell's native spiking vocabulary could, in theory, serve as a prosthetic cerebellar output, feeding corrective motor signals downstream to the deep cerebellar nuclei that ultimately control movement. That application remains years away from any animal trial, let alone human use, but the proof of concept now exists in a dish on a lab bench in Evanston, Illinois.
Against This Article's Thesis
Mouse cerebellar slices in a petri dish are not living brains in skulls, and that single fact threatens to collapse the entire "writing to the brain" narrative. Every neuroscientist knows this caveat, and it is worth stating at full strength: the gap between stimulating neurons in vitro and safely interfacing with a functioning brain is enormous. Immune response, long-term biocompatibility, signal-to-noise degradation in vivo, encapsulation failure, thermal effects of continuous operation, and the regulatory pathway through FDA approval all stand between this lab demonstration and any clinical reality. Neuralink's approach, descendant of the decades-old Utah microelectrode array, works in living humans right now, and Noland Arbaugh browses the internet with his thoughts using a device that required brain surgery, a skull-mounted connector, and months of patient calibration. Printed MoS2 neurons have zero human data, zero in-vivo animal data, and zero evidence of long-term stability in biological environments. Calling this a "speaker" for the brain is technically defensible but practically premature by a decade or more.
What This Analysis Does Not Prove
Energy-per-spike estimates for the printed neurons are inferred from voltage and current data in the paper, not directly reported by the authors as a benchmark; the actual energy consumption in a deployed biointerface setting would depend on drive circuitry, interconnects, and encapsulation overhead that do not yet exist. The Purkinje stimulation was performed in acute brain slices, meaning tissue was alive but not connected to a functioning nervous system; the neurons' response to printed-circuit stimulation does not demonstrate that downstream motor circuits would receive or correctly interpret the signal. Cycle stability exceeding one million operations was measured under controlled lab conditions, not in a saline or cerebrospinal fluid environment where corrosion and biofouling would degrade the MoS2 memristors. The 20 kHz frequency ceiling, while impressive, exceeds the physiological firing range of most neurons (typically under 200 Hz); the practical utility of this headroom for biointerface applications is unclear. No data on multi-channel integration, signal multiplexing, or the minimum electrode density needed for meaningful neural writing was reported.
The Bottom Line
If you work in BCIs, neural prosthetics, or neuromorphic engineering, here is what to watch. Track whether Hersam's lab publishes in-vivo animal data; the jump from brain slice to living rodent is where most neural interface technologies stall or die, and no timeline for that experiment has been disclosed. For biotech investors, the combination of printable manufacturing, neuron-native waveforms, and demonstrated tissue interfacing represents a genuinely new category; the comparable moment was when flexible electronics moved from academic curiosity to commercial wearable sensors around 2015, and printed neural interfaces could follow a similar ten-year trajectory from lab to clinic. For neuroscience researchers, the key immediate value is not the biointerface potential at all but the spiking-complexity toolkit: six distinct neuron behaviors in a printed, tunable, reproducible circuit that could replace expensive and fragile biological preparations in electrophysiology labs, potentially standardizing how researchers study neural coding. For everyone else, the signal buried in this Nature Nanotechnology paper is structural: the BCI field's 50-year assumption that neural interfaces must be fabricated in semiconductor cleanrooms using rigid silicon or metal electrodes just acquired its first serious counterexample, printed on plastic, with an ink jet, for a fraction of the cost.
Sources
- Hadke, S.S., Klingler, C.N., Brown, S.T., et al. (April 2026). "Printed MoS2 memristive nanosheet networks for spiking neurons with multi-order complexity." Nature Nanotechnology. DOI: 10.1038/s41565-026-02149-6
- Dewan, S. (April 2026). "Scientists invent artificial neurons that talk to real brain cells, paving way to better brain implants." Live Science. Live Science
- Izhikevich, E.M. (2004). "Which model to use for cortical spiking neurons?" IEEE Transactions on Neural Networks. Izhikevich taxonomy
- Intel Corporation. Loihi 2 neuromorphic processor: architecture and benchmarks. Intel Neuromorphic Computing
- Northwestern University McCormick School of Engineering. Hersam Research Group. mccormick.northwestern.edu