Amorphous Oxide Semiconductor (AOS) Selectors

Updated 5 July 2026

AOS selectors are amorphous oxide materials that enable disorder-tolerant electron conduction through mixed or single-cation transport networks.
They are engineered by tuning orbital overlaps and oxygen stoichiometry to optimize conduction and achieve higher Hall mobilities.
In monolithic 3D DRAM, BEOL-compatible oxide selectors serve as routing elements that reduce parasitic capacitance and boost density scaling.

Searching arXiv for the cited papers and topic context. Amorphous oxide semiconductor (AOS) selectors are selector elements or selector-relevant amorphous transport media whose operation is governed by the electronic structure of disordered metal-oxide or metal-oxy-nitride networks and, in architectural implementations, by the use of BEOL-compatible oxide select transistors to gate access between local interconnects and shared routing resources. In the cited literature, the topic spans two connected levels. At the materials level, amorphous IGZO and amorphous ZnON are analyzed to determine how cation- $s$ -derived conduction paths, orbital overlap, and oxygen stoichiometry shape electron transport in amorphous networks. At the system level, oxide selectors are deployed as routing devices in monolithic 3D DRAM, where they are used to manage routing congestion, hybrid bonding constraints, parasitics, disturb, and sensing margin (Srivastava et al., 2018, Lee et al., 12 Mar 2026, Kang et al., 2011).

1. Scope and selector roles

The cited work uses “selector” in two distinct but related senses. In monolithic 3D DRAM, the selector is an explicit routing/select transistor placed at the cell-array-to-strap interface, physically above the cell array, where it gates access from local bitlines or wordlines to strapped or global routing. In the amorphous-semiconductor studies, selector relevance is indirect: the papers do not study threshold switching directly, but they analyze atomistic and electronic-structure features that would condition the behavior of an amorphous semiconductor used as the active selector medium or transport layer (Lee et al., 12 Mar 2026, Srivastava et al., 2018).

A critical distinction in the 3D DRAM study is between the cell access transistor inside each 1T1C DRAM cell and the routing/select transistor used in the proposed BL Selector + Strap architecture. The selector does not replace the DRAM cell access transistor. Rather, it is an additional transistor in the line-routing path. The same paper also uses “AOS” in a second role, as the cell access transistor channel, and compares 3D (Si) and 3D (AOS) access-transistor cases. This dual usage is central to the terminology of AOS selectors in vertically integrated memory (Lee et al., 12 Mar 2026).

At the materials level, two further scope limitations are explicit. The comparison between a-IGZO and a-ZnON is “not about selector devices per se,” and the atomistic study of a-IGZO and oxygen deficiency “does not” simulate threshold switching, filament formation, dynamic snapback, or selector current-voltage curves. The selector relevance of these papers therefore lies in their account of disorder-tolerant conduction, mixed-cation versus single-cation transport topology, and defect chemistry rather than in direct two-terminal selector measurements (Srivastava et al., 2018, Kang et al., 2011).

2. Disorder-tolerant conduction in amorphous oxides

In amorphous oxide semiconductors, the conduction-band minimum is described as unusually tolerant to amorphous disorder because it is dominated by metal $s$ -states. For a-IGZO, the atomistic study reports that the electronic structure around the CBM is little affected by disorder and that band-tail states are few generated for the CBM. The paper attributes this to the fact that the CBM is characterized almost only by cation- $s$ orbitals, which are spherical and nondirectional, whereas the valence-band maximum is dominated by O- $p$ states and is strongly disorder-sensitive and localized (Kang et al., 2011).

The a-IGZO versus a-ZnON comparison refines this picture. In both amorphous systems, the conduction-band edge is still dominated by metal $s$ -states, but the transport network differs sharply. In a-ZnON, the conduction-band minimum is primarily built from Zn-4s orbitals. In a-IGZO, the conduction-band minimum contains substantial contributions from In-5s, Zn-4s, and Ga-4s, and the authors explicitly argue against the oversimplified view that transport in a-IGZO occurs only through overlapping In-5s orbitals. The valence-band edge differs by anion chemistry: O-2p dominates in a-IGZO, while N-2p dominates in a-ZnON (Srivastava et al., 2018).

This distinction is fundamental for selector-oriented interpretation. A mixed-cation AOS such as a-IGZO presents a chemically mixed percolation network, whereas a single-cation / multiple-anion system such as a-ZnON presents a single-cation transport backbone. This suggests that the electronic uniformity of the conduction network may differ even when both systems retain the standard AOS feature of disorder-tolerant cation- $s$ conduction. A plausible implication is that the simpler backbone of a-ZnON should be more electronically uniform than the mixed-cation network of a-IGZO, although the cited paper does not measure selector switching directly (Srivastava et al., 2018).

3. Transport topology, orbital overlap, and mobility descriptors

The a-IGZO and a-ZnON study makes the transport picture explicitly structural by introducing the orbital overlap integral (OI) between neighboring metal $s$ -orbitals. The authors use Mulliken’s formulation for Slater-type orbitals and numerically evaluate the integrals in Mathematica. For each pair type inside a 4 Å interaction sphere around each metal cation, they compute the number of cation-cation pairs, the average separation $R$ , the pairwise overlap integral at that average distance, and the volume-normalized quantity $\text{Pairs} \times \text{OI} / \text{Volume}$ . The resulting total normalized sum is the practical transport descriptor, sometimes described as the total normalized orbital overlap integral. Effective masses are extracted from pseudo-band curvature using

$\frac{1}{m^*}=\frac{1}{\hbar^2}\left(\frac{\partial^2E}{\partial k^2}\right),$

with a parabola fitted to the conduction-band minimum over a small range of $s$ 0 $s$ 1 around $s$ 2 along $s$ 3-X and then averaged across three orthogonal directions (Srivastava et al., 2018).

The paper’s quantitative comparison is summarized below.

Material	CBM character at $s$ 4	Transport descriptor
a-ZnON-21	Zn-s 55.53%, N-s 14.79%, O-s 13.01%	total normalized OI 0.06231; $s$ 5
a-ZnON-11	Zn-s 49.90%, N-s 21.09%, O-s 7.24%	total normalized OI 0.06693; $s$ 6
a-IGZO-1114	In-s 29.10%, O-s 22.78%, Zn-s 20.11%, Ga-s 14.92%	total normalized OI 0.02966; $s$ 7
a-IGZO-2217	In-s 35.76%, O-s 22.23%, Ga-s 16.29%, Zn-s 11.71%	total normalized OI 0.02669; $s$ 8

The structural interpretation is equally important. In a-IGZO-1114, the major pair contributions include Zn-Zn, In-In, Ga-Ga, In-Zn, In-Ga, and Zn-Ga, with the dominant normalized contributions coming from mixed-cation pairs such as In-Zn, Zn-Ga, and In-Ga. In a-IGZO-2217, the same mixed-cation pattern remains. In a-ZnON-21 and a-ZnON-11, the transport channel is reduced to Zn-Zn overlaps only. The total normalized OI in ZnON is therefore more than 2× larger than in IGZO, and the authors explicitly connect this with the experimentally reported much higher Hall mobilities in a-ZnON. For context, the paper states that a-IGZO Hall mobilities are typically up to about 15 cm $s$ 9/V·s, while a-ZnON has reported Hall mobilities exceeding 200 cm $s$ 0/V·s. It also stresses that direct quantitative comparison between computed OI or effective mass and measured Hall mobility is difficult because mobility also depends on scattering and carrier concentration (Srivastava et al., 2018).

The study further notes an earlier criterion from Orita et al. that for transport in oxide semiconductors the metal- $s$ 1 orbital overlap should exceed 0.4 and the fraction of transport-forming cations should exceed a 20% percolation threshold. The calculated pair OIs in this work are all below 0.4, roughly 0.13–0.30, and the authors attribute the discrepancy to differences in how the overlap is computed. Even so, the percolative interpretation remains important: in IGZO there are not enough same-cation links to support a single-species pathway, so transport must be mixed. This suggests that mixed-cation oxide networks may naturally produce a broader distribution of local barriers and transport bottlenecks than a single-cation network (Srivastava et al., 2018).

4. Oxygen deficiency, localization, and defect physics in a-IGZO

The first-principles study of amorphous InGaZnO $s$ 2 distinguishes sharply between a metastable oxygen-vacancy-like configuration and a relaxed O-deficient amorphous structure obtained by removing an oxygen atom and then repeating simulated annealing. Its key conceptual claim is that O-deficient a-IGZO should not simply be described as a point oxygen vacancy defect. After annealing and structural reconstruction, oxygen deficiency is better understood as a change in local substructures and coordination motifs rather than as a fixed localized vacancy center (Kang et al., 2011).

The wavefunction-localization analysis is based on the expansion

$s$ 3

the site weights

$s$ 4

and the normalized inverse participation ratio

$s$ 5

Using this metric, the paper reports that the IPR values near the CBM in stoichiometric a-IGZO are similar to crystalline IGZO, confirming that localized band-tail states are surprisingly absent or very weak around the CBM. It also reports that the VBM is strongly localized and exhibits pronounced valence-band tail states, reflecting the disorder sensitivity of O- $s$ 6 states (Kang et al., 2011).

For moderately O-deficient a-IGZO—one missing oxygen in the supercell after full annealing—the paper finds no localized donor-center near the CBM. Free electron carriers can be generated without creation of donor-level in the O-deficient amorphous oxide, and the CBM wavefunction remains delocalized. The reported energetic data are specific: metastable O-vacancy-like structures in a-IGZO have formation enthalpies from 0.77 eV to 4.61 eV; after additional simulated annealing with one missing oxygen, the total energy becomes lower by 0.45 eV than the most stable metastable $s$ 7 structure; and the resulting neutral shallow-donor-like O-deficiency state has a formation enthalpy of 0.32 eV. By contrast, oxygen-vacancy formation enthalpies in crystalline IGZO are 4.14 \sim 4.64 eV (Kang et al., 2011).

For severe O-deficiency—two missing oxygens in the supercell—the behavior changes qualitatively. Structural fluctuations become stronger, weak or dangling bonds form, an asymmetric InO $s$ 8 motif with an isolated dangling bond appears, and a deep donor level in the gap is created that traps two electrons. The authors state that repeated simulated annealing runs could not avoid deep-level formation in this heavily deficient regime. For selector interpretation, this suggests that oxygen stoichiometry is likely the key knob separating disorder-tolerant electron conduction from deep-trap-dominated degradation in IGZO-based amorphous switching layers. That interpretation remains an inference, because the paper itself does not model threshold switching or selector current-voltage nonlinearity (Kang et al., 2011).

5. BEOL oxide selectors in monolithic 3D DRAM

In monolithic 3D DRAM, AOS selectors appear not as speculative materials but as explicit routing devices in a bitline strap architecture with selectors. The architecture uses vertical BL (VBL) organization. Bitlines are grouped into 8 BLs per strap, wordlines are grouped into 16 WLs per strap, the BL strap is routed in Metal 2 (M2), and the selectors are integrated on top of the cell array in BEOL-compatible form. The paper states explicitly that “select transistors are integrated on top of the cell to alleviate routing congestion,” and identifies the selector as an Indium-Gallium-Oxide (IGO [11]) device placed in the routing path between local cell-array BLs/WLs and the strapped or global routing connected to bonded periphery (Lee et al., 12 Mar 2026).

The architectural motivation is obtained by comparing four routing schemes: direct BLSA connection, BL strapping, core MUX, and BL Selector + Strap. Direct connection and core multiplexing require extremely tight hybrid bonding pitch: 0.26 µm HCB pitch for Si and 0.22 µm HCB pitch for AOS. Strapping alone relaxes pitch to 0.75 µm (Si) and 0.62 µm (AOS), but the paper states that simple strapping “substantially increases BL parasitic capacitance (CBL), thereby degrading the sense margin.” The selector+strap scheme preserves the relaxed pitch while isolating inactive branches, reducing effective capacitive loading on the active bitline. The paper reports effective $s$ 9 = 6.6 fF including bonding parasitics, compared to 20 fF for D1b, and states that the select transistor enables the inactive BL to float at a refresh potential, “effectively mitigating the FBE and off-state leakage by decoupling the cell from the global BL” (Lee et al., 12 Mar 2026).

The selector device assumptions are specific. In Fig. 6, the BEOL selector is a DG transistor with gate length 50 nm and gate width 70 nm. Its performance is given as $p$ 0 for $p$ 1, subthreshold slope approximately 60 mV/dec, and on-off ratio $p$ 2. The selector is described as BEOL compatible, integrated above the cell, and used because of its performance and BEOL compatibility (Lee et al., 12 Mar 2026).

At the full-array level, the paper treats the selector+strap scheme as “essential.” The optimized design achieves 2.6 Gb/mm $p$ 3, representing ~6× density scaling over D1b 2D DRAM, with 137 layers with Si access transistors or 87 layers with AOS. The same work reports sense margin = 130 mV (Si), 189 mV (AOS) for selector+strap, compared with 54 mV for D1b. It reports a nominal row cycle time $p$ 4 of 10.5 ns, compared to 21.3 ns in D1b, and read/write energies of 6.26/1.57 fJ (Si) and 5.38/1.35 fJ (AOS), corresponding to a 60% reduction in read/write energy. These results are presented as a consequence of the selector’s ability to relax bonding pitch, contain bitline capacitance, preserve sensing margin, and reduce disturb and leakage exposure (Lee et al., 12 Mar 2026).

6. Design implications, misconceptions, and unresolved issues

Three misconceptions are corrected directly by the cited literature. First, transport in a-IGZO should not be described as an In-5s-only channel; the conduction path is a mixed-cation $p$ 5-orbital network involving In, Zn, and Ga. Second, oxygen deficiency in amorphous IGZO should not automatically be reduced to a localized oxygen-vacancy picture; after structural reconstruction it may generate free electrons without a discrete donor-like localized level near the CBM. Third, in monolithic 3D DRAM the AOS selector is not the same device as the 1T1C access transistor; it is a complementary routing and isolation element used in the BL Selector + Strap architecture (Srivastava et al., 2018, Kang et al., 2011, Lee et al., 12 Mar 2026).

The design heuristics that follow are partly direct and partly inferential. Directly, the cited materials work shows that a-ZnON conducts through a Zn-4s-only network, whereas a-IGZO conducts through a mixed network of overlapping metal $p$ 6-orbitals; a-ZnON has higher total normalized orbital overlap and lower effective mass than a-IGZO; and the simpler single-cation network is qualitatively consistent with much higher reported Hall mobility (Srivastava et al., 2018). This suggests that, if one is engineering an amorphous semiconductor as the active selector medium or transport layer, a single-cation transport backbone may offer better spatial uniformity of electronic coupling than a mixed-cation oxide. It also suggests that the total normalized orbital overlap integral can serve as a practical, structure-derived descriptor for qualitative mobility screening across many amorphous compositions and structures (Srivastava et al., 2018).

At the same time, the cited work is explicit about its limits. The a-IGZO versus a-ZnON comparison does not study threshold switching, two-terminal selector geometry, OFF-state leakage, current-voltage nonlinearity, trap-assisted hopping under high field, Poole-Frenkel or Schottky-limited transport, filament-free switching dynamics, Joule heating, transient recovery, endurance, selector/current-limiter integration with memory, or variability of threshold voltage and current. The atomistic oxygen-deficiency study likewise does not address threshold switching directly. The monolithic 3D DRAM study focuses on TCAD/SPICE feasibility and routing architecture and does not provide direct yield or reliability data for the selector itself (Srivastava et al., 2018, Kang et al., 2011, Lee et al., 12 Mar 2026).

Taken together, the cited papers establish a layered view of AOS selectors. At the network level, amorphous transport depends on whether the conduction path is a chemically mixed percolation network or a chemically simpler single-cation network. At the defect level, oxygen stoichiometry controls whether a-IGZO remains in a disorder-tolerant, weakly trapped electron-transport regime or crosses into deep-trap formation associated with distorted In-centered motifs. At the architectural level, compact BEOL oxide selectors can become indispensable routing elements that make monolithic 3D DRAM feasible under realistic bonding and parasitic constraints. The selector-specific consequence for abrupt threshold switching remains unresolved in these works, but the materials and architecture design rules they provide are precise: examine whether the amorphous conduction network is chemically simple, whether the total normalized metal- $p$ 7 orbital overlap is high or low, and whether oxygen stoichiometry is controlled tightly enough to avoid the transition from benign carrier generation to deep localized defect formation (Srivastava et al., 2018, Kang et al., 2011, Lee et al., 12 Mar 2026).