Today's article comes from the journal of Memories - Materials, Devices, Circuits and Systems. The authors are Shih et al., from Imec, in Belgium. In this paper they're showcasing a new type of computer memory. It works by storing binary information as ordered chains of two different nanoparticle types inside billions of liquid-filled nanocapillaries.
DOI: 10.1016/j.memori.2026.100143
If you look at a photo of a computer from the 50s or 60s you're probably going to see a giant machine (refrigerator size or larger), with banks of metal cabinets, rows of switches and indicator lights, control consoles, card readers, line printers, cooling vents and bundles of thick cables running between everything. And mounted somewhere near the main processor will be what looks like a reel-to-reel or a movie projector. Two or more big wheels of plastic or metal with long strands of thin polymer strips wrapping around them and passing from one to another. But that's not film, no. It's magnetic tape.
Long before we had SSDs, magnetic disk drives, CDs or floppy disks, tape was used to store information. Business records, scientific measurements, financial transactions or anything else a computer might need to preserve. You see the tape itself is coated in a layer of magnetizable particles, typically iron oxide. And the system works by moving that tape past an electromagnetic write head, which changes the magnetic orientation of tiny regions in that coating. Those orientations are arranged into coded sequences that represent binary data. So when the computer needs to store a file, for example, it winds the tape forward while the write head records that file as a sequence of magnetic transitions. Then when that file needs to be retrieved, it winds through the tape until it reaches the correct location and uses a read head to detect those transitions and reconstruct the original bits.
This kind of system is simple, durable and quite cheap. But it's also big, cumbersome and slow.
Why? Because the system may have to wind through a long stretch of tape (and unspool or spool much of those wheels) before reaching the requested record. So over time, as newer, smaller, faster storage mechanisms came on the scene, more and more machines began abandoning tape in favor of those alternatives. But that doesn't mean that tape disappeared entirely. In fact, quite the opposite. Today tape is, perhaps surprisingly, still used for enterprise backups, long-term scientific archives, cloud provider infrequent-access storage and more. Why? Because while SSDs might be better at rapid random access, magnetic hard drives might be easier to update, and pretty much every other option might be more convenient to access, tape is extremely economical. At least when enormous amounts of data must be retained and rarely opened. It costs very little per terabyte, it draws essentially no power while sitting on a shelf, it can be physically isolated and airgapped from attackers, and it lasts 30 years or more. So if you're working on any kind of super-long-term project and you've got to store petabytes of information, you might still find yourself reaching for one of these reels. Or, more accurately today, for a tape cartridge (the modern incarnation of these systems).
That being said, the industry is actively trying to change that. To find something that can approach tape's low cost (and enormous capacity) while providing faster access, greater parallelism and less mechanical handling. Ideally, that technology would still retain data without continuous power, and pack immense quantities of it into a small physical area. But unlike tape, it would be able to retrieve individual objects without winding through a long sequential medium, and be able to perform updates just as conveniently.
But that's easier said than done. Density, access speed, manufacturing cost, retention and reliability pull storage designs in different directions.
One of the more recent contenders, and one of the first to mount a credible challenge across tape's territory, is called 3D NAND. It works by stacking flash-memory cells vertically instead of arranging them across a flat surface. Manufacturers deposit many alternating material layers, etch extremely deep vertical channels through the stack and form long strings of memory cells around those channels. Each cell stores one or more bits by trapping different quantities of electrical charge, and the device reads those charge states electronically without any moving parts.
But there's an issue: increasing the capacity of 3D NAND requires manufacturers to add more layers, etch deeper and narrower holes, maintain uniform electrical behavior across the entire stack and connect all of those structures together. Those kinds of high-aspect-ratio processes are not just technically demanding, but require enormous capital investment. And they become harder to scale economically as the stacks grow taller. So while 3D NAND is extraordinarily successful in some areas, it is still not inexpensive enough (at scale) to make tape unnecessary for the largest cold-storage or archival workloads. Though it did, arguably, come close.
What the field still needs is a storage technology that fills the gap between online solid-state storage and offline tape: something with extremely high bit density, nonvolatile retention, moderate throughput, millisecond-scale access, and an architecture that does not require a separate complex access device for every stored bit. And in today's paper, the authors think they've found exactly that. It's called ultra-high bit density liquid-state colloidal memory, and it works by storing binary information as ordered chains of two different nanoparticle types inside billions of liquid-filled nanocapillaries. And actually, the authors didn't just find it. They invented it a few years ago. This same team, at Imec, has been painstakingly developing and refining this technology since at least 2020. Trying to get it from a vague patent-application spec (with a lot of missing pieces) to a flushed out and usable idea. And in this paper they're making the case that this concept just moved a little bit closer to the finish line. That it is now physically plausible, that suitable particles and access mechanisms can in principle be engineered, and that a mature device could provide nearline storage with densities far beyond conventional HDDs and potentially beyond projected 3D NAND. On today's episode we'll learn how exactly this technology works, and then we'll walk through the particular contribution of this most recent paper. Let's dive in.
At the highest level, the authors' invention (which we'll call colloidal memory for short) consists of a silicon substrate containing a dense two-dimensional array of vertical nanocapillaries. Above those capillaries is a sealed liquid reservoir containing two kinds of suspended nanoparticles. One particle type represents a 0 and the other represents a 1. Each capillary is only slightly wider than the particles themselves, so particles entering it cannot move around one another. Their order is therefore preserved, and the vertical chain inside one capillary becomes a physical bit string. So unlike 3D NAND, where the bits are charge states, this system stores each bit in the identity of a discrete physical particle. And unlike tape or an HDD, it contains no macroscopic moving reel, platter or mechanical read head. Its moving components are just nanoparticles and liquid confined inside tiny channels.
Writing to colloidal memory begins with a pair of electrodes positioned near the opening of each capillary. When those electrodes receive an alternating voltage, they generate a nonuniform electric field. A polarizable particle in that field then experiences dielectrophoresis, meaning that it is pushed either toward or away from the strongest part of the field depending on its electrical properties and the frequency being applied. The authors call the two proposed classes "protagonistic" and "antagonistic" particles.
The controller can therefore choose which bit to write simply by changing the frequency, attracting one particle from the mixed reservoir and excluding the other. And repeating that operation builds an ordered sequence one particle at a time.
Getting the particles into a capillary is only part of the problem. Once written, they must remain trapped for years without drifting back into the reservoir under Brownian motion. The proposed design, therefore, places a geometric constriction and an electrostatic energy barrier near the entrance. During storage, that barrier is high enough to prevent particles from escaping. And during writing or erasing, electrode pairs reshape the local energy landscape, capture the selected particle, move it across the constriction and then restore the barrier behind it. Then, to move the complete chain during reading, the device can use electro-osmotic flow. A direct electric field acts on mobile ions near the charged capillary walls, those ions drag the liquid with them, and the flowing liquid transports the particle stack along the channel without requiring a conventional mechanical pump.
But what about reading? Well, reading relies on the fact that the two particle types interact differently with an electric field. As each particle passes a sensing electrode, it changes the capacitance between the electrodes by a characteristic amount. The circuitry detects those changes and reconstructs the binary sequence.
But, as I said, the authors didn't actually introduce this technology here. The debut was a few years ago. Today's paper is all about determining whether or not it could actually work.
To find out, the authors worked backward from the proposed operating mechanism to determine what physical properties the two particles would need. They used a conventional core-shell electrical model and then a more sophisticated numerical model incorporating electrostatics, hydrodynamics and electrokinetic effects. Their results point toward a highly surface-charged dielectric particle, such as functionalized polystyrene, for the protagonistic state and a metallic particle with little or no surface charge for the antagonistic state. After that, they modeled how long it would take to capture a particle, how strong the retention barrier would need to be, how large the readout signal might be and how quickly particles could be translated through a capillary. Under their assumptions, they estimate a random-access latency of about 1.3 milliseconds and aggregate throughput of about 310 megabytes per second when operating 16,000 capillaries in parallel. That places the concept somewhere between 3D NAND and HDDs in performance. They also estimated the potential density. A device using 25-nanometer particles, 10-micrometer-deep capillaries and a 50-nanometer capillary pitch could theoretically store 160 gigabits per square millimeter. Shrinking the particle diameter and pitch further could raise that estimate to roughly 740.
Then, finally, it was on to fabrication. The team built a real preliminary test chip containing 80-nanometer-wide, 300-nanometer-deep capillaries and exposed it to fluorescent 25-nanometer polystyrene particles. At 10 megahertz, the particles experienced negative dielectrophoresis and were not captured. And at 100 kilohertz, their response switched to positive dielectrophoresis and they accumulated inside the capillaries. That is an important proof of concept because it shows that the electric force can overcome random Brownian motion at the required particle scale. But it is not yet a working memory. The experiment did not selectively write both particle types into controlled bit strings, retain them for years, read them electronically or operate a large integrated array. They did make another step in that direction, but there's still a lot left to do.
Overall, this technology is still at the exploratory and proof-of-principle stage, several steps before even a laboratory memory prototype. The next milestone would be an integrated device that repeatedly selects both particle types, writes known binary sequences, preserves their order, reads those sequences electrically, and then erases or rewrites them with an acceptably low error rate. After that would come larger parallel arrays, error-correction systems, endurance testing and long-duration retention experiments. The researchers must also still figure out how to solve liquid packaging on CMOS, particle aggregation, solvent stability, corrosion, contamination, readout sensitivity and manufacturing uniformity across billions of capillaries. If those problems can be solved, colloidal memory could (in theory) eventually occupy a new nearline tier between SSDs and tape. But right now, this work demonstrates just a credible physical mechanism (and to be fair a potentially extraordinary density) but not a storage product that is anywhere close to commercialization.