Microelectrode Arrays (MEAs) Explained: A Complete Guide

A microelectrode array (MEA) is a grid of tiny electrodes, each typically tens of micrometers wide, built into the floor of a culture dish or the surface of a chip. Neurons grown on it can be recorded at many sites at once and stimulated with small electrical pulses, making it the main two-way interface for cultured neurons and brain organoids.

Key takeaways

An MEA “listens” to neurons extracellularly: each electrode picks up voltage changes in the fluid around nearby cells, not the inside of any single cell.
Passive MEAs have tens to hundreds of electrodes wired out individually; CMOS high-density MEAs integrate electronics under the electrodes, reaching 26,400 electrodes at a 17.5 micrometer pitch in one well-documented design.
Recordings split into two bands: fast spikes from individual neurons and slower local field potentials from populations.
Stimulation normally uses biphasic, charge-balanced pulses to protect both the electrode and the tissue.
Flat arrays only sample the bottom of a 3D brain organoid, which is why shell-shaped and mesh electronics are an active research area.

What is a microelectrode array?

At its simplest, a microelectrode array is a set of conductive pads, each connected to an amplifier, sitting under a layer of living cells. When a neuron near a pad fires an action potential, ions rush across its membrane. That current flows through the surrounding salt solution and produces a small, brief voltage change at the electrode. The amplifier picks it up, a digitizer samples it, and software turns it into data.

Because the electrode sits outside the cell, this is called extracellular electrophysiology. It is less detailed than patch clamping, which pokes a glass pipette into or onto one cell, but it has two decisive advantages. It does not damage the cells, so the same culture can be recorded for weeks or months. And it scales: one chip can follow hundreds or thousands of neurons at once.

The same electrodes can also push current into the tissue. That makes an MEA a two-way channel: you can read what a network is doing and write patterned input back into it. This two-way property is what makes the microelectrode array the foundation of organoid intelligence and biocomputing research, where living neurons are treated as a computing substrate. For the bigger picture of that field, see our guide to what organoid intelligence is.

A short history

MEAs are older than many people assume. Thomas and colleagues described a miniature microelectrode array for monitoring cultured cells in Experimental Cell Research in 1972, and Pine reported recording action potentials from cultured neurons with extracellular microcircuit electrodes in the Journal of Neuroscience Methods in 1980. Through the following decades, glass-substrate MEAs with a few dozen electrodes became a standard lab tool for studying network activity, drug effects and plasticity in dissociated cultures and brain slices.

The big change came when engineers began building arrays directly on CMOS silicon, the same process used for computer chips. That moved the amplifiers from a rack beside the microscope to a few micrometers below the cells.

How do passive and CMOS high-density MEAs differ?

The key engineering question for any MEA is how to get signals from many electrodes out to amplifiers without the wiring becoming impossible.

Passive MEAs

In a passive MEA, each electrode is a metal pad (often titanium nitride, platinum, gold or indium tin oxide) connected by its own metal track to a contact pad at the edge of the chip. External amplifiers connect to those pads. Every electrode needs its own track, so the number of electrodes is limited by how many tracks you can squeeze in. Passive arrays commonly have tens to a few hundred electrodes, spaced on the order of a hundred micrometers or more apart.

Passive MEAs are simple, robust and often transparent, which makes it easy to image the cells through the glass. Their limitation is spatial sampling. With wide spacing, most neurons are not close to any electrode, and a single electrode cannot tell you much about where a signal came from.

CMOS high-density MEAs

A high-density microelectrode array (HD-MEA) solves the wiring problem by putting active circuitry under each electrode or group of electrodes. Amplifiers, filters, switch matrices and analog-to-digital converters live in the chip itself.

A well-documented example comes from the Hierlemann group at ETH Zurich. Ballini and colleagues (2014, IEEE Journal of Solid-State Circuits) described a single CMOS chip with 26,400 platinum electrodes over a 3.85 by 2.10 mm sensing area, at a pitch of 17.5 micrometers. Any user-selected subset of electrodes can be routed to 1,024 on-chip readout channels, and the chip includes 32 stimulation units. The authors reported noise of 2.4 microvolts rms in the action-potential band and 5.4 microvolts rms in the local field potential band, with 10-bit digitization at 20,000 samples per second. Müller and colleagues (2015, Lab on a Chip) showed the platform recording from over a thousand cells at once, at subcellular, cellular and network levels, including following signals along individual axons.

This class of chip is what Kagan and colleagues used for DishBrain, the system that let cultured neurons play a simplified game of Pong (Neuron, 2022). Their methods describe a commercial CMOS array that reads up to 1,024 of its 26,400 electrodes at 20,000 samples per second. We explain that experiment in how neurons learn in a dish.

Comparing the main MEA types

Feature	Passive planar MEA	CMOS high-density MEA	3D shell or mesh interface
Typical electrode count	Tens to a few hundred	Thousands to tens of thousands (with a subset read at once)	Tens of electrodes in published research devices
Electrode spacing	Roughly 100 µm or more	Below 20 µm in leading designs	Distributed across a 3D surface or volume
Electronics	External amplifiers	On-chip amplifiers, multiplexing, digitization	Usually external, research prototypes
Strength	Simple, transparent, robust, inexpensive	Single-neuron and axonal resolution, flexible electrode selection	Samples the sides or interior of 3D tissue
Main limitation	Sparse sampling	Large data volumes, opaque silicon, fixed flat geometry	Fabrication complexity, early-stage, few channels
Common use	Drug screening, network activity	Detailed network mapping, closed-loop experiments	Brain organoid and spheroid research

What do electrode size, spacing and sampling rate change?

Three parameters shape what an MEA can see.

Electrode size. Smaller electrodes pick up signals from fewer neurons and give finer spatial detail, but they have higher impedance, which means more thermal noise and less capacity to deliver stimulation charge. Larger electrodes average over more tissue. Engineers counter the impedance problem with rough or porous coatings (such as platinum black or conductive polymers) that increase the effective surface area without increasing the footprint.

Spacing (pitch). When electrodes are closer together than the size of a neuron’s electrical footprint, one neuron’s spike appears on several neighboring electrodes with different amplitudes. That spatial pattern is a fingerprint, which greatly helps spike sorting. It also lets you track an action potential as it travels along an axon.

Sampling rate. An extracellular spike lasts on the order of a millisecond. To capture its shape, systems sample far faster than that, typically in the range of 10,000 to 30,000 samples per second per channel. The CMOS designs above run at 20,000 samples per second. Local field potentials are slower and can be downsampled heavily without losing information.

A useful mental model: electrode size and spacing decide which cells you can separate, and sampling rate decides how much detail you get about each event.

What do extracellular signals look like?

If you plot raw voltage from one electrode, you see a noisy trace with occasional sharp downward deflections riding on slower wobbles. Filtering separates these into two bands, a convention that the Ballini chip builds into its noise specifications (action-potential band 300 Hz to 10 kHz; local field potential band 1 Hz to 300 Hz).

Spikes

Spikes are the extracellular footprint of action potentials from neurons close to the electrode. They are brief (about a millisecond), usually negative-going at the cell body, and small, typically tens to a few hundred microvolts, much smaller than the roughly 100 millivolt swing measured inside the cell. Amplitude falls off quickly with distance, so an electrode mostly “hears” neurons within a short radius.

Spikes are the currency of most decoding work. Count them in time bins and you get firing rates; look at their exact times and you get spike trains, which carry information about timing and correlation between cells.

Local field potentials

The local field potential (LFP) is the slower part of the signal. Buzsáki, Anastassiou and Koch explained in Nature Reviews Neuroscience (2012) that extracellular fields arise from the superposition of all the ionic currents in a region, with synaptic currents a major contributor, alongside other sources such as spikes and intrinsic membrane currents. The LFP therefore reflects coordinated input and activity across a population rather than the output of a single cell.

In cultures and organoids, LFP-band signals are often dominated by network bursts: episodes in which much of the network fires together, followed by quiet periods. Trujillo and colleagues (2019, Cell Stem Cell) used MEAs to follow human cortical organoids over several months and reported periodic oscillatory network events that later became more irregular, with some features resembling preterm human EEG.

Signal-to-noise ratio

Signal-to-noise ratio (SNR) decides whether you can detect and separate spikes at all. Noise comes from the electrode-electrolyte interface (thermal noise grows with impedance), the amplifier, biological background from many distant neurons, and environmental interference such as mains hum. Key ways to improve SNR:

Lower electrode impedance, through coatings that increase effective surface area.
Low-noise amplification close to the electrode, which is a core advantage of CMOS arrays.
Tight cell-electrode contact, because any gap of fluid between cell and electrode shunts signal away.
Shielding and grounding, to keep line noise and handling artifacts out.
Spatial averaging and template matching on dense arrays, which use the fact that one spike appears on several electrodes.

A common detection rule sets a threshold at a multiple of the estimated noise level. Kagan and colleagues, for example, report setting their spike threshold at six standard deviations above background noise for their activity scans.

How does an MEA stimulate neurons?

Recording is listening. Stimulation is talking. An MEA stimulates by passing current (or imposing a voltage) through selected electrodes, which changes the voltage across the membranes of nearby neurons. If the change is large enough, a neuron fires.

Biphasic, charge-balanced pulses

The workhorse waveform is the biphasic pulse: a brief phase of one polarity followed by a phase of the opposite polarity, often each lasting tens to hundreds of microseconds. In a charge-balanced pulse, the two phases carry equal charge, so no net charge is left behind.

Why does that matter? Merrill, Bikson and Jefferys reviewed the electrochemistry in the Journal of Neuroscience Methods (2005). Injecting charge into tissue drives reactions at the electrode surface. As long as charge stays within the electrode’s reversible capacity, it moves back and forth without permanent chemical change. Push it further, or leave net charge in place, and irreversible reactions can corrode the electrode, change local pH and generate products that harm cells. Their review sets out how waveform shape, charge per phase and charge density shape both effectiveness and safety.

In practice, experimenters choose:

Which electrodes to stimulate (one site, a pattern, or a moving sequence).
Pulse shape, usually biphasic and charge-balanced, with the leading phase chosen to excite most effectively.
Amplitude and duration, kept within safe charge limits for the electrode material.
Timing, from single pulses to trains at a set frequency, or patterns that encode information.

Stimulation artifacts

A stimulation pulse is enormous compared with a spike, and it can saturate amplifiers on nearby electrodes for a short time. Systems handle this by “blanking” (ignoring) samples right after each pulse, using amplifiers that recover quickly, and subtracting artifact templates. DishBrain’s software, for example, ignored all 1,024 channels for a configurable number of samples after each stimulation command so that the pulse itself was not mistaken for neural activity.

Encoding information as stimulation

To use a culture or organoid for computation, you have to translate data into stimulation, a step called encoding. Common schemes include place coding (which electrode is stimulated represents a value), rate coding (pulse frequency represents a value) and temporal patterns. DishBrain combined place and rate coding to signal the ball’s position. Decoding is the reverse step: turning recorded activity back into numbers. Both are covered in our glossary entry on neural encoding and decoding. When encoding and decoding run continuously, with the output shaping the next input, the result is closed-loop stimulation.

How does the spike sorting pipeline work?

Raw MEA data is a set of voltage traces. Most scientific and computational uses need something else: a list of which neuron fired when. Spike sorting bridges the gap. A typical pipeline has these stages:

Filtering. A band-pass filter (for example from a few hundred hertz to several kilohertz) removes the slow LFP and high-frequency noise, leaving the spike band.
Common noise removal. Signals shared across all electrodes, such as interference, are estimated and subtracted.
Detection. Candidate spikes are found where the voltage crosses a threshold, usually a multiple of a robust noise estimate.
Feature extraction. Each candidate is summarized, for example by waveform shape, principal components or wavelet coefficients, and, on dense arrays, by its amplitude pattern across neighboring electrodes.
Clustering. Candidates with similar features are grouped, on the assumption that each cluster corresponds to one neuron. Quiroga and colleagues (2004, Neural Computation) introduced a widely used unsupervised approach combining wavelets and superparamagnetic clustering.
Template matching and drift correction. Modern sorters build a template for each putative neuron and match it against the data, handling overlapping spikes and slow movement of the tissue relative to the electrodes. Kilosort4 (Pachitariu and colleagues, Nature Methods, 2024) is a recent example of this family.
Curation and quality control. Clusters are checked for refractory period violations, stability and isolation, often with some manual review.

The output is a set of “units” with spike times, which can feed analysis, visualization or a decoder. Results are probabilistic. Two neurons with similar waveforms may merge, and one neuron may be split, so careful work reports quality metrics alongside results.

How much data does an MEA produce?

High-density arrays generate data quickly. Here is a worked example with simple assumptions:

1,024 channels recorded simultaneously
20,000 samples per second per channel
2 bytes per sample (a common storage size for 10 to 16-bit samples)

That gives 1,024 × 20,000 × 2 = about 41 megabytes per second, roughly 2.5 gigabytes per minute and about 147 gigabytes per hour of raw voltage. A multi-week experiment that records continuously quickly reaches tens of terabytes per array.

This is why practical systems lean on a few strategies:

On-line spike detection, storing spike times and short waveform snippets instead of every raw sample.
Selective raw recording, keeping full-bandwidth data only around key events or for limited periods.
Compression, since neural data compresses reasonably well.
Standard file formats, so that data, metadata and stimulation logs stay together. Neurodata Without Borders (NWB) is the leading open standard in neurophysiology.

For closed-loop work, there is an extra constraint: latency. The time from a spike occurring to the system responding with new stimulation must be short enough to matter to the neurons, ideally a few milliseconds to tens of milliseconds. DishBrain counted spikes in 10 millisecond windows to update its game.

What are the challenges of recording from 3D brain organoids?

MEAs were designed for flat cultures: a monolayer of cells spread across the electrodes. A brain organoid is different. It is a 3D ball of tissue that can grow to a few millimeters across, grown from induced pluripotent stem cells through a process we describe in how brain organoids are grown.

Only the bottom layer is sampled

Put a sphere on a flat array and only a small patch touches the surface. Extracellular signals fall off fast with distance, so electrodes mostly detect neurons in the lowest layers of tissue. Activity in the interior and on the top and sides is largely invisible. Huang and colleagues put it plainly in Science Advances (2022): conventional MEAs, designed for monolayer cultures, offer limited recording contact area restricted to the bottom of 3D organoids.

Contact, movement and health

Other practical issues compound the geometry problem:

Uneven contact. Organoids are not perfectly round, and they may attach at a few points rather than across a flat face.
Movement and drift. Organoids can shift on the surface over days, which changes which neurons each electrode sees and complicates spike sorting.
Nutrient and oxygen supply. Large organoids can develop poorly supplied cores. Approaches such as microfluidics and organoid vascularization aim to improve this, which also matters for long recordings.
Variability. No two organoids are identical, so electrode coverage and signal quality vary from sample to sample.

Some researchers slice organoids or let them flatten partially onto the array to increase contact. Others plate them on HD-MEAs and accept that they sample one surface in great detail. In the Brainoware study (Cai and colleagues, Nature Electronics, 2023), an organoid sat on a high-density MEA and served as the computing “reservoir”; coverage of the work noted the flat, rigid geometry of the arrays as a limitation. We explain that work in reservoir computing with brain organoids.

What new interfaces are emerging for 3D tissue?

Several groups are developing electrodes that wrap around or grow through 3D tissue. All are research-stage, and published devices usually have far fewer channels than CMOS planar arrays, but they point to where the field is heading.

3D frameworks for spheroids

Park and colleagues (2021, Science Advances) introduced microfabricated 3D frameworks that act as compliant interfaces to cortical spheroids and assembloids, with electrical, optical, chemical and thermal functions. They reported studying how coordinated bursting events spread across the surface of a spheroid, and how bridging tissue formed between pairs of spheroids, a model related to the assembloid concept.

Self-folding shell MEAs

Huang and colleagues (2022, Science Advances), a team including Johns Hopkins researchers, drew inspiration from EEG caps. They built miniature, optically transparent shells from self-folding polymer leaflets carrying conductive-polymer-coated metal electrodes. The leaflets fold around an organoid, and the authors reported recording from organoids 400 to 600 micrometers in size for up to four weeks, including responses to glutamate. The point is coverage: electrodes touch the sides and top of the organoid, not just the bottom.

Stretchable mesh nanoelectronics

Le Floch and colleagues (2022, Advanced Materials) took a different route. They designed stretchable, “tissue-like” mesh electronics with mechanical properties matched to developing organoids, and let the tissue fold the mesh into itself during growth. The result distributes electrodes through the 3D volume. They reported that the mesh did not interrupt organoid development, maintained stable contact over long periods, and captured single-cell action potentials emerging from early development.

What these approaches trade off

Approach	What it gains	What it costs
Planar HD-MEA	Thousands of electrodes, mature electronics, fine pitch	Only samples the contact surface
Shell MEA	Coverage of the outer surface, transparency for imaging	Fewer electrodes, complex fabrication, organoid size must match
Mesh electronics grown into tissue	Access to the interior over long periods	Device must be integrated early, hard to scale channel count, specialized fabrication

The Frontiers in Science roadmap for organoid intelligence by Smirnova and colleagues (2023) identifies 3D interfaces with organoids as one of the key technical requirements for the field, alongside larger, more mature organoids and better data infrastructure.

Why do microelectrode arrays matter for biocomputing?

Every wetware computing experiment depends on the interface. The neurons cannot receive a task, and we cannot read their response, except through electrodes (or, in some labs, light-based methods). That means the MEA sets hard limits on:

Bandwidth: how many independent input and output channels a living system can have.
Resolution: whether we work with single neurons or population averages.
Speed: how fast a closed loop can run.
Duration: how long one preparation can be studied without damage.

Better arrays make better experiments possible. Denser electrodes help separate individual neurons and their synaptic plasticity over time. 3D interfaces bring more of the organoid into view. Lower-latency electronics tighten the closed loop. Comparisons with silicon approaches such as neuromorphic computing and spiking neural networks also depend on measuring the living system accurately in the first place. And the energy argument for using neurons at all, which we examine in biocomputing vs silicon, must eventually account for the power drawn by the recording hardware itself.

The software side matters just as much. Choosing electrodes, scheduling stimulation, running spike detection in real time, logging every pulse and storing results in open formats are all problems that sit between the chip and the science. This is the problem we’re working on at Organoid Programming: a protocol SDK and closed-loop runtime that make MEA experiments reproducible and shareable.

Conclusion

A microelectrode array is the ear and the voice of any neuron-based computing system. Planar passive arrays gave researchers their first long-term view of network activity decades ago. CMOS high-density MEAs now record from thousands of electrodes at single-neuron resolution, generating tens of megabytes per second and enabling real-time closed loops. The frontier is 3D: shell-shaped and mesh electronics that reach beyond the bottom layer of a brain organoid.

Understanding the trade-offs, from electrode size and noise to charge-balanced stimulation and spike sorting, is the first step in judging any claim about organoid intelligence. For where the field may go next, read the future of biocomputing, and for the questions it raises, see the ethics of organoid intelligence.

Frequently asked questions

What is a microelectrode array (MEA)?

A microelectrode array is a chip or dish whose surface carries a grid of tiny electrodes, typically tens of micrometers across. Neurons or tissue placed on it can be recorded extracellularly at many sites at once, and the same electrodes can deliver small electrical pulses. MEAs are the standard interface for long-term recording and stimulation of cultured neurons, brain slices and brain organoids.

What is the difference between a passive MEA and a CMOS high-density MEA?

A passive MEA routes each electrode through its own wire to external amplifiers, which limits it to tens or a few hundred electrodes spaced far apart. A CMOS high-density MEA builds amplifiers, multiplexers and digitizers into the silicon chip under the electrodes. That allows thousands of electrodes at pitches below 20 micrometers, fine enough to follow single neurons and even their axons.

What signals does an MEA record?

MEAs record extracellular voltages. After filtering, the high-frequency band (roughly 300 Hz upward) contains spikes, the brief signatures of action potentials from nearby neurons, typically tens to a few hundred microvolts in size. The low-frequency band (below about 300 Hz) contains local field potentials, which mainly reflect summed synaptic currents from many cells around the electrode.

How do MEAs stimulate neurons safely?

Most protocols use biphasic, charge-balanced pulses: a short pulse of one polarity followed immediately by an equal and opposite pulse, so that no net charge is left at the electrode. This limits electrochemical reactions that can damage the electrode and the tissue. Researchers also keep charge per pulse and charge density within limits appropriate for the electrode material.

What is spike sorting and why does it matter for MEAs?

Spike sorting is the process of detecting spikes in extracellular recordings and assigning each one to the neuron that most likely produced it, based on waveform shape and, on dense arrays, its spatial footprint across electrodes. Without it, an electrode’s activity is a mix of several cells. Sorting turns raw voltage into per-neuron spike trains that can be analyzed or decoded.

Can a flat MEA record from a whole brain organoid?

No. A flat MEA only contacts the cells at the bottom of a three-dimensional organoid, so most of the tissue is out of range of the electrodes. Organoids may also contact the surface unevenly. Researchers are developing 3D interfaces, such as self-folding shell electrodes and flexible mesh electronics, to sample activity from the surface and interior of organoids.

How much data does a high-density MEA produce?

A lot. As a worked example, a CMOS array recording 1,024 channels at 20,000 samples per second, stored at 2 bytes per sample, produces about 41 megabytes per second, or roughly 147 gigabytes per hour. Long experiments therefore rely on on-line spike detection, compression and careful data management with standard formats.

Microelectrode arrays explained: how we listen to and talk with neurons