Atomristors that switch in reverse, found by screening 2,940 monolayers

What the work claims

Atomristors are metal-insulator-metal devices in which the active layer is a single two-dimensional monolayer, typically a transition-metal dichalcogenide or hexagonal boron nitride, and switching happens when metal atoms from the electrode adsorb into native vacancies in the layer, introducing conducting states. Since the first demonstration in monolayer MoS2 in 2018, the materials pool had barely grown beyond a handful of compounds. The authors built a high-throughput pipeline combining density functional theory, machine-learning interatomic potentials for molecular dynamics, and non-equilibrium Green's function quantum transport, and screened the 2DMatPedia database for new candidates.¹

Two claims emerge. First, scale: of 2,940 theoretically exfoliable monolayers, 61 survive electronic and chemical filters; 26 show the required adsorption-induced phase transition; and 17 remain after molecular-dynamics stability checks at 300 K and 700 K, 8 in the conventional class and 9 in the new one. Second, novelty: the new class, exemplified by BiOCl, is "inverted". Its native vacancy donates electrons that delocalize and push the Fermi level into the conduction band, so the fresh device sits in the low-resistance state; gold adsorption re-localizes those carriers, re-opens a band gap of 2.78 eV, and drives the device to the high-resistance state. Conventional devices such as GaS need a forming step, a first high-voltage pulse to create the conductive path; inverted devices do not, which the authors argue suits reset-dominated workloads and complementary circuit pairing.¹

How it works

The screening funnel is worth stating precisely, because the filtering choices are themselves design information. From 2,940 top-down exfoliable monolayers in 2DMatPedia, only insulators with a band gap above 1.5 eV pass (910), then exfoliation and decomposition energies below 100 meV per atom (694), then hazardous elements, f-block elements and unit cells above six atoms are excluded, leaving 64, of which 61 retain their gap under the authors' recalculated electronic structure. Into these they introduce every elemental monovacancy, compute formation energies, and keep the lowest-energy native defect for each material. Gold is the chosen switching species because gold electrodes dominate the experimental atomristor literature. Of 59 defective supercells examined, 26 show an electronic phase transition on gold adsorption: 10 conventional, insulator-to-metal (class 1), and 16 inverted, metal-to-insulator (class 2). Molecular dynamics with a benchmarked machine-learning force field (Orb-v3) at 300 K and 700 K then removes 9 candidates with structurally unstable defects, leaving 8 class 1 materials, including GaS, GaSe and PtS2, and 9 class 2 materials, including BiOCl, BiOBr, ZrNBr and HfNCl.¹

Device-level quantum transport simulations on two representatives, vertical Au-GaS-Au and Au-BiOCl-Au stacks with a deliberately imposed 5 Angstrom electrode-monolayer separation, give concrete numbers. For GaS, gold adsorption drops the resistance by a factor of 3.8 x 10^2 at 0.1 V bias; for BiOCl, gold adsorption raises the resistance by about 0.6 x 10^2, roughly sixtyfold. Current-voltage traces from 0 to 0.6 V give a low-to-high-state current ratio of about 2 x 10^2 for GaS at 0.6 V and 2.4 x 10^1 for BiOCl, with the BiOCl ratio rising to 8.4 x 10^1 at the opposite polarity. The inverted window is real but smaller than the conventional one.¹

The kinetic analysis is where the paper is most honest. At the relaxed, equilibrium electrode-monolayer separation, gold atoms spontaneously migrate into vacancies within 10-picosecond molecular-dynamics runs at 300 K, typically within the first 2 picoseconds, for GaS in two crystal phases, GaSe, PtS2 and BiOCl. In other words the relaxed contact is not stable: the device programs itself. Increasing the electrode separation by 20 percent suppresses the spontaneous adsorption and enables voltage-triggered switching, but the same increase raises the gold migration barrier and so demands higher switching voltages. There is a direct trade between memory window and switching energy, controlled by one knob, the electrode-to-monolayer distance. The authors also quantify retention through nudged-elastic-band reverse barriers: benchmarks WS2 and MoS2 give 1.04 and 0.92 eV, and among the new candidates HfNCl and one phase of HfNBr exceed 1 eV, corresponding to an idealized isolated-state residence time above 2 x 10^4 seconds. They flag clearly that these are single-path Arrhenius estimates, comparative indicators rather than retention predictions.¹

Where a skeptic should push

Nothing here has been fabricated. Every number above is a simulation output, and several known gaps separate these idealized stacks from devices. The imposed 5 Angstrom separation that produces the quoted windows is precisely the quantity a real process cannot control that well, and the paper's own supplementary analysis shows the window collapsing as the electrode approaches the monolayer, because metal-induced gap states bridge the high-resistance state. At realistic contact distances the quoted ratios should be treated as upper bounds, and the field's own history of wildly scattered measured on-off ratios for identical materials, which the authors cite, is consistent with exactly this sensitivity.

The deeper load-bearing assumption is that a single, freely diffusing gold atom is a good write mechanism. It is elegant at the atomic scale and brutal at the circuit scale: writing means driving electrode metal into the active layer, which is electromigration, the same failure mode that ends CMOS interconnects. No endurance data of any kind exists here, and the introduction itself notes the atomristor field's poor endurance and low fabrication yield as standing obstacles. Second, the inverted class's headline advantage, forming-free operation, comes packaged with its headline risk: at equilibrium spacing these devices spontaneously adsorb gold and would, on the authors' own simulations, leave fabrication already programmed. Whether that is a feature (self-initializing state) or a defect (uncontrolled programming) is an open experimental question. Third, the screening excludes anything with a unit cell above six atoms, hazardous elements and f-electron chemistry, so the 17-candidate list is a map of what was easy to compute, not necessarily of what is best to build. Weight this as high-quality computational discovery with credible device physics, zero validation.

Inverted switching and the array memory fabric

The obvious read for a microelectrode array is density: a monolayer-thick non-volatile element stackable in the back end of a CMOS readout chip, which is the tier where per-electrode state wants to live. Calibration trims, per-site amplifier configuration, stimulation amplitude history and reference patterns for spike-sorting verification are all data that change rarely and must survive power loss, and today they either live off-chip or consume SRAM area that competes with the photodiode-and-electrode real estate an array actually sells. A two-terminal element under a quarter of a nanometre thick that holds its state with no refresh is the right shape for that job, and the class 2 result adds something the field's memory elements rarely offer: the as-fabricated state is the conducting state, so no forming pulse is ever needed. Forming is exactly the operation a tissue-adjacent chip cannot tolerate, because it is a high-voltage pulse delivered through electronics millimetres from a living preparation.

The paper's genuinely transferable finding, though, is about interfaces, and it cuts the other way. The memory window dies when a conductor gets within Angstroms of the active layer, because metal-induced gap states short out the insulating state. Now consider where an on-array memory element would sit: on one side, metal interconnects; on the other, an electrolyte with a tissue whose extracellular space is itself a weakly conductive, ionically screened medium pressing against the passivation. The physics that erodes the atomristor window at a gold contact is the same physics the electrode community fights as leakage through thin passivation and as partial screening of the recorded field. The design rule the paper hands over, that separation distance trades window against switching voltage, becomes an encapsulation-thickness trade for any array that co-locates memory and electrodes: every extra nanometre of hermetic isolation that protects the tissue-facing side also weakens the coupling the memory needs, and the window you measured on a test wafer is not the window you will have after years in warm saline.

The endurance picture also inverts surprisingly in the array's favor. A write mechanism based on electrode-metal migration will never survive the 10^6-cycle life of a digital cache, which is why these elements are poor candidates for streaming spike buffers. But the per-electrode state an array actually needs to store, calibration and configuration, changes on timescales of days to months. For that workload a low-endurance, forming-free, monolayer-thin store with seconds-long Arrhenius-limited retention is not a deficient memory; it is close to correctly specified. The honest residual risks are the ones the paper names: as-fabricated self-programming at equilibrium spacing, and retention that is a single-path estimate rather than a measurement.

The bottom line

Established, in silico: a 2,940-material screen converges on 17 thermally stable atomristor candidates; a class 2 family including BiOCl switches from natively conducting to insulating on gold adsorption, opposite to every atomristor demonstrated before; quantum transport at imposed separation gives resistance windows of 3.8 x 10^2 (GaS) and about 60 (BiOCl) at 0.1 V; and electrode-to-monolayer separation is the central device knob, trading spontaneous adsorption, memory window and switching voltage against each other. Not established: a single fabricated device, any endurance, or measured retention; the quoted windows are upper bounds at an interface spacing no process currently guarantees. What would confirm the story is a fabricated GaS or BiOCl vertical cell whose measured window tracks the predicted spacing dependence and whose retention approaches the barrier-implied timescales; what would break it is contact-controlled fabrication collapsing the windows below one order of magnitude, which prior literature on identical materials suggests is a live risk. For array instrumentation the durable takeaway is twofold: forming-free monolayer memory is a plausible near-pixel storage tier for rarely changed state, and the interface physics that governs its window is the same interface physics the acquisition chain already has to engineer against warm saline.