# IGZO Neuromorphic Systems: Evidence and Materials Technical Report

Version 3.1 · 12 October 2026 · IGZO Evidence Atlas

## Scope and evidence rule

This research note compares materials, device, and array claims in the supplied project documents with primary academic sources. It reports no new device measurement. Peer-reviewed experiment, circuit/network simulation, patent claim, and project target are distinct evidence classes. Results from different devices cannot be combined into a synthetic platform specification.

## 1. Materials and physics

For amorphous In–Ga–Zn–O (IGZO), preparation route, oxygen chemistry, electrode interface, film thickness, and thermal history can affect switching. An I–V curve alone does not establish a unique mechanism. The device stack, electrode materials, microscopy/chemical analysis, and controls should be reported together. Martins and colleagues reported eight states and retention up to 10^5 seconds for particular solution-processed IGZO devices under stated conditions [1]. Those values are not guaranteed for another process or architecture.

Naqi and colleagues' fabricated 10×10 IGZO array is a primary device/array example [2]. The approximately 500 cycles and >10^4 seconds of retention are a different evidence class from the reported CNN/SNN recognition figures, which are associated with electrical and network simulation. They do not demonstrate a completed sensor-to-actuator hardware inference loop for this project.

## 2. Measurement vocabulary

- **State window:** G_high/G_low alone is insufficient. Report each state's distribution and overlap with adjacent states.
- **Retention:** Declare read voltage, temperature, sampling schedule, pass tolerance, and the model used to extrapolate accelerated tests.
- **Endurance:** Define a SET/RESET cycle, failure rule, and censoring. A median over surviving cells is insufficient.
- **Variability:** Separate within-cell, cell-to-cell, row/column, wafer, and batch variation. A thousand cells in one batch are not a thousand independent batches.
- **Yield:** Distinguish functional-cell yield, usable-array yield, and calibrated-system yield.
- **Energy:** For writing, E_write = ∫ V(t)I(t)dt. Include all pulses, verification reads, and retries until successful programming.

Instrument bandwidth, probe connections, compliance setting, thermal conditions, random sampling, and raw-data retention should be preregistered. JCGM/NIST uncertainty guidance provides a methodological basis for traceable units and uncertainty estimates [3].

## 3. Fabrication and acceptance gate

Scenario A sources finished devices. B1 fabricates film/electrode/patterning in-house; B2 outsources fabrication while retaining independent incoming QC. All three routes require one acceptance schema: stack and geometry, process revision, raw I–V traces, SET/RESET distributions, read noise, retention, endurance, read disturb, batch yield, and rejects. Progress to 32×32 requires prespecified state separation and error budgets. A 64×64 array is conditional, not automatic.

## 4. Claim corrections

The <1 ms, <1 W, 300×, and 10^6-cycle phrases in supplied drafts lack project-specific raw measurements and remain targets or hypotheses. The up-to-82% SET and 54% READ energy comparison concerns a HfO₂/CMOS 3T1R circuit, not an IGZO material result [4]. Patent EP 3 588 594 B1 concerns a particular structure; it does not prove universal write-once security [5]. Confidential-agreement text and party details were not transferred to the public site.

## 5. Minimum publication package

Retain sample/batch identifiers, process versions, mask and geometry, instrument calibration metadata, every raw I–V and pulse trace, failed samples, prespecified exclusions, analysis code, and independent replication instructions. Charts without provenance and sample counts are not sufficient evidence. All 3D images on the site are conceptual illustrations.

## References

1. Martins et al., *Journal of Materials Chemistry C* (2022), https://doi.org/10.1039/D1TC05465A
2. Naqi et al., *Materials Today Nano* 25, 100441 (2024), https://doi.org/10.1016/j.mtnano.2023.100441
3. NIST Measurement Uncertainty, https://www.nist.gov/itl/sed/topic-areas/measurement-uncertainty
4. Das et al., preprint (2023), https://arxiv.org/abs/2305.16418
5. EP3588594B1, https://patents.google.com/patent/EP3588594B1/en

## Portfolios 1–2 audit addendum (v3.1)

The two PDF portfolios contain 13 embedded files, with individual presentations repeated in compiled volumes. Repetition is not independent measurement. The 10^6-cycle statement in patents and drafts cannot be assigned to the project device. Wang et al. report a different 1-kb Pt/IGZO/SiNₓ/Ta array with >10^6 cycles and >10^6 s retention. Device stacks and test conditions must be matched before transferring any result. Source: https://www.nature.com/articles/s41467-026-72740-x