rudymartin/rsct-memristor-benchmark

RSCT Memristor Benchmark

A multi-device memristor benchmark for evaluating representation--substrate compatibility.

The RSCT Memristor Benchmark is a curated collection of memristor device measurements spanning multiple material families, measurement techniques, and device architectures. It is designed for evaluating whether computational representations (neural network weights, conductance maps, synaptic update rules) are compatible with physical memristor substrates through the Representation--Substrate Compatibility Testing (RSCT) framework.

This benchmark is the substrate-facing counterpart to GeoRSCT, which evaluates representation--solver compatibility in the geospatial domain.

The companion Python package, processing scripts, and CI pipeline are maintained at https://github.com/NextShiftConsulting/rsct-memristor.

Theoretical Foundation

This benchmark is an artifact of Representation--Solver Compatibility Theory (RSCT).

The underlying theory is introduced in:

Rudolph A. Martin. Intelligence as Representation-Solver Compatibility: A General Theory of Representation-Dependent Reasoning. Preprint, March 3, 2026. Zenodo

RSCT argues that observed performance depends on the relationship among the problem, the encoding, and the solver:

D = D(P, E, S)

In the substrate domain, the question becomes: given a computational representation trained on a digital solver, can it be faithfully mapped onto a physical memristive substrate without unacceptable degradation? The RSCT Memristor Benchmark provides the device-level measurements needed to answer that question through 4-gate certification.

Dataset Summary

The benchmark includes 8 registered datasets spanning 7 device families, all processed into Parquet tables (805,288 total rows).

ID	Device Family	Material System	Measurements	Rows	Source DOI
milano_quantum_2025	Quantum point contact	Ag/Ag2S	I-V, conductance quantization, metrology	3,112	10.5281/zenodo.16788655
siox_gradual_2021	Oxide RRAM	SiOx	SET/RESET I-V sweeps	1,004	10.5281/zenodo.5762184
dang_fatigue_2025	Chalcogenide flexible	Ag2Se	I-V endurance (1000 cycles), STDP, noise	208,064	10.5281/zenodo.17501455
polymeric_stdp_2022	Organic polymer	PEDOT:PSS	PPF, STDP timing	85	10.5281/zenodo.12685756
rivera_sierra_2025_pulses	Nanofluidic	Iontronic	Pulse train, PPF, I-V sweep	50,488	10.5281/zenodo.17159696
bismuth_perovskite_2026	Perovskite	Bi halide	I-V cycles, forming, retention, endurance	92,916	10.5281/zenodo.17799630
rf_sparameters_2026	RF memristor	--	S-parameters (70 kHz -- 160 GHz)	9,600	10.5281/zenodo.20186567
memriki_logic_2025	Iontronic (simulation)	--	I-V, gate logic, dynamics, repeatability	440,019	10.5281/zenodo.14924500

Data Sources and Provenance

All source data was obtained from Zenodo, an open-access research data repository operated by CERN. Each dataset was deposited by the original research team and is linked to a peer-reviewed publication.

Dataset	Original Authors	Institution	Publication	License
milano_quantum_2025	Brun-Picard et al.	LNE (France), PTB (Germany), INRIM (Italy)	Nature Nanotechnology, 2025 (10.1038/s41565-025-02037-5)	CC-BY-4.0
siox_gradual_2021	Barmpatsalos, Mehonic	UCL (UK)	Standalone EPSRC dataset	CC-BY-4.0
dang_fatigue_2025	Dang et al.	Tsinghua University (China)	Nature Electronics, 2026 (10.1038/s41928-025-01554-4)	CC-BY-4.0
polymeric_stdp_2022	Panthi et al.	Czech Academy of Sciences	Materials Advances, 2024 (10.1039/D4MA00399C)	CC-BY-4.0
rivera_sierra_2025_pulses	Rivera-Sierra, Bisquert	Universitat Jaume I / ITQ-UPV-CSIC (Spain)	--	CC-BY-4.0
bismuth_perovskite_2026	Kim, Rivera-Sierra, Mengesha, Iniguez, Bisquert	ITQ-UPV-CSIC (Spain), URV (Spain)	Advanced Materials Technologies, 2026	CC-BY-4.0
rf_sparameters_2026	--	--	--	CC-BY-4.0
memriki_logic_2025	--	Utrecht University (Netherlands)	arXiv 2503.13386v2	MIT

No modifications were made to the measured data. Processing consists solely of format conversion (txt/xlsx/csv/s2p/jld2 to Parquet), column renaming for consistency, reshaping multi-column layouts into tidy tables, and subsampling of very large simulation datasets (memriki_logic: 390M rows subsampled to 160K). The original raw files are preserved in the companion GitHub repository under data/<dataset_id>/raw/ (gitignored; downloadable from the Zenodo DOIs above).

Parquet Tables

Each dataset contains one or more Parquet tables under data/<dataset_id>/:

milano_quantum_2025

• iv_experimental.parquet -- Voltage-current sweeps from Figures 2c/2d (2,759 rows)
• conductance_traces.parquet -- G/G0 conductance time series from Figures 3a-3f (341 rows)
• metrology.parquet -- NMI inter-laboratory metrology comparison from Figures 4a/4b (12 rows)

siox_gradual_2021

• iv_set_reset.parquet -- SET and RESET I-V sweeps with current (A), voltage (V), and resistance (ohm) columns (1,004 rows)

dang_fatigue_2025

• iv_endurance_1000cycles.parquet -- 1,000-cycle I-V endurance sweeps: voltage, current, cycle number (200,000 rows)
• iv_comparison.parquet -- Dual sweep comparison (596 rows)
• conductance_modulation.parquet -- Multi-condition conductance modulation (6,068 rows)
• stdp_timing.parquet -- Spike-timing-dependent plasticity pre/post conductance (999 rows)
• fatigue_spike_response.parquet -- Multi-channel spike response over cycles (200 rows)
• fatigue_conductance.parquet -- Multi-channel conductance evolution (201 rows)

polymeric_stdp_2022

• ppf.parquet -- Paired-pulse facilitation index vs inter-pulse interval (45 rows)
• stdp.parquet -- STDP weight change vs timing at 10 ms and 100 ms intervals (40 rows)

rivera_sierra_2025_pulses

• iv_pulse_train.parquet -- I-V pulse train responses at 1--5 V amplitudes (1,330 rows)
• iv_sweep.parquet -- I-V sweeps at +/-2 V to +/-5 V ranges (48,070 rows)
• ppf_pulse_response.parquet -- Paired-pulse facilitation response curves from Figs 6--7 (1,088 rows)

bismuth_perovskite_2026

• iv_cycles.parquet -- 30-cycle I-V switching curves for I-rich and Br-rich compositions (30,000 rows)
• forming_iv.parquet -- Forming I-V curves from 50 independent cells per composition (50,000 rows)
• set_reset_voltage.parquet -- SET/RESET voltage distributions (58 rows)
• resistance_states.parquet -- HRS, LRS, and ON/OFF ratio statistics (58 rows)
• endurance.parquet -- HRS/LRS evolution over 1000 endurance cycles (1,999 rows)
• retention.parquet -- HRS/LRS retention stability over time (2,002 rows)
• iv_scan_rate.parquet -- Scan-rate-dependent I-V curves (8,799 rows)

rf_sparameters_2026

• s_parameters.parquet -- S11/S21/S12/S22 (real+imaginary) across 70 kHz to 160 GHz for 2 memristor devices in multiple resistance states, plus calibration standards (9,600 rows)

memriki_logic_2025

Simulation data from ACME.jl circuit simulation (not physical measurement). Original JLD2 files contain 424M timesteps; uniformly stride-decimated to ~20K points per file. See data/memriki_logic_2025/SAMPLING.md for full disclosure of sampling method, aliasing analysis, and impact assessment.

• iv_simulation.parquet -- Single memristor I-V at 0.4, 40, and 4000 Hz with reconstructed input voltage (60,002 rows; from 60M)
• gate_logic.parquet -- NAND and XOR logic gate input/output voltage traces (40,002 rows; from 400K)
• dynamics.parquet -- Driven Shinriki oscillator at 3 frequency/resistance configs (60,001 rows; from 330M)
• repeatability.parquet -- 100-loop gate repeatability for AND, OR, NAND, XOR with multi-stage cascades (160,008 rows; from 81M)
• combined_gates.parquet -- Sequential NAND-based AND and OR gate cascades, 3 stages each (120,006 rows; from 1.2M)

What This Benchmark Is For

Use the RSCT Memristor Benchmark to study:

1. Whether a computational representation can be mapped onto a specific memristor substrate.
2. Whether device nonidealities (cycle-to-cycle variability, drift, asymmetric SET/RESET) degrade representation fidelity below acceptable thresholds.
3. Whether conductance update linearity and dynamic range support the precision required by a given neural network architecture.
4. Whether STDP timing windows in physical devices match the requirements of bio-inspired learning rules.
5. Whether substrate-level measurements support or refute admissibility claims for specific device--representation pairings.

This benchmark should not be used as a device ranking or as evidence that any specific device "works" for neuromorphic computing without additional system-level validation.

RSCT Certification

Each dataset can be evaluated through the RSCT 4-gate certification process. The simplex decomposes the measurement space as R + S_sup + N = 1, where R = relevant signal, S_sup = superfluous content, and N = noise. The substrate certificate implements a simplified version of the canonical gate pipeline:

1. Gate 1 (Integrity): N < 0.5 -- Noise below integrity threshold?
2. Gate 2 (Consensus): S_sup <= 0.6 -- Superfluous content not dominant? (Simplified; canonical uses multi-source coherence.)
3. Gate 3 (Admissibility): kappa_coupling >= 0.5 -- Substrate coupling above admissibility threshold? (Simplified; canonical uses Oobleck dynamic scaling.)
4. Gate 4 (Grounding): R >= 0.3 -- Sufficient relevant signal for grounding?

Install the companion package:

pip install rsct-memristor

from rsct_memristor.certificates.schema import MemristorCertificate

cert = MemristorCertificate(
    dataset_id="dang_fatigue_2025",
    device_family="chalcogenide_flexible",
    R=0.65,
    S_sup=0.25,
    N=0.10,
    kappa_coupling=0.65 * (1 - 0.10),  # R*(1-N) = 0.585
)
cert.apply_gates()
print(cert.decision)  # "substrate_admissible"

Getting Started

import pandas as pd

# Load a specific table
df = pd.read_parquet("data/dang_fatigue_2025/iv_endurance_1000cycles.parquet")
print(f"{len(df)} rows, {df.columns.tolist()}")

# Or use the companion package
from rsct_memristor.datasets.loader import load_dataset_local
tables = load_dataset_local("dang_fatigue_2025")
for name, df in tables.items():
    print(f"{name}: {df.shape}")

Considerations for Using the Data

Cross-Device Comparability

Datasets from different memristor devices use different materials, instruments, units, and measurement domains. Raw I-V curves from different devices are not directly comparable. Users should extract device-agnostic substrate metrics (conductance window, update linearity, cycle-to-cycle variability, relaxation time constants) from each dataset independently and compare derived metrics, not raw measurements.

Known Limitations

1. Device diversity is limited. Eight datasets across six device families provide breadth, but do not cover all memristor technologies (e.g., phase-change, ferroelectric, and spintronic devices are not represented).
2. Measurement conditions vary. Temperature, sweep rate, voltage range, and instrumentation differ across datasets. These are documented in the original publications but not normalized in the Parquet tables.
3. No system-level validation. This benchmark contains device-level measurements. Crossbar array behavior, sneak-path effects, peripheral circuit overhead, and system-level power consumption are out of scope.
4. Simulation data is included. The memriki_logic_2025 dataset is from circuit simulation (ACME.jl), not physical measurement. It is included for completeness but should not be treated as equivalent to experimental data.
5. Small sample sizes for some datasets. The polymeric STDP dataset (85 total rows) provides limited statistical power. Results from small datasets should be interpreted with appropriate uncertainty.

Personal and Sensitive Information

This dataset contains no individual-level or personally identifiable data. All data consists of electrical measurements from memristor devices performed in laboratory settings.

Additional Information

Dataset Curators

The RSCT Memristor Benchmark was curated by Next Shift Consulting as part of the RSCT (Representation--Solver Compatibility Theory) research program. All raw data was produced by the original research teams listed in the Data Sources table above.

Licensing

The RSCT Memristor Benchmark is released under CC-BY-4.0.

All source datasets are licensed CC-BY-4.0 by their original authors, except memriki_logic_2025 which is MIT-licensed. CC-BY-4.0 is adopted as the binding minimum across the collection.

When using or redistributing this benchmark, preserve attribution to the original dataset authors via the Zenodo DOIs listed in the Data Sources table.

Changelog

• v0.2.0 (2026-06-23): All 8 datasets processed to Parquet (805,288 rows). Added rivera_sierra, bismuth_perovskite, rf_sparameters, memriki_logic (full JLD2 extraction via Julia). Corrected RSCT field names to canonical (R, S_sup, N, kappa_coupling). Updated gate descriptions. Sampling disclosure for memriki_logic.
• v0.1.0 (2026-06-23): Initial release. 4 of 8 datasets processed to Parquet (212,265 rows). Companion package at v0.1.0.

Citation

If you use the RSCT framing or representation--solver compatibility terminology, please cite the theory preprint:

@misc{martin2026rsct,
  title        = {Intelligence as Representation-Solver Compatibility: A General Theory of Representation-Dependent Reasoning},
  author       = {Martin, Rudolph A.},
  year         = {2026},
  note         = {Preprint. Zenodo DOI: 10.5281/zenodo.18854651; SSRN: 6339299}
}

If you use this benchmark dataset, please cite:

@dataset{rsctmemristor2026,
  title        = {RSCT Memristor Benchmark: Multi-Device Substrate Compatibility Evaluation},
  author       = {Martin, Rudolph A.},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/rudymartin/rsct-memristor-benchmark},
  version      = {0.1.0},
  note         = {8 memristor datasets across 7 device families; 525,274 processed rows; CC-BY-4.0}
}

Please also cite the original dataset authors whose Zenodo DOIs are listed in the Data Sources table.