MemDeviceBench

MemDeviceBench is an open Python toolkit and metadata standard for reproducible benchmarking of 2-terminal, 3-terminal, 4-terminal, and multi-terminal memory/adaptive electronic devices.

The project is deliberately not named after ECRAM, RRAM, or any other single device family. Device identity is represented by independent tag axes, so one dataset can carry several physically different descriptors without forcing them into one flat category.

Status: v0.2.0 alpha. The validated numerical core currently targets pulse-by-pulse conductance-update traces. DC I–V, transfer/output curves, polarization/PUND, transient waveforms, noise, and impedance are present in the taxonomy and roadmap, but are not all implemented as analysis profiles yet.

The central idea: topology first, tags by axis

RRAM, PRAM/PCM, FRAM, FeFET, and ECRAM are technology or device-concept labels. TFT, OECT, capacitor, resistor, and transistor describe platforms. 2T, 3T, and 4T describe the electrically accessible terminals of the measured DUT. A physically meaningful record therefore keeps these axes separate.

Example DUT	Technology tags	Accessible topology	Platform tags	Possible mechanism tags
oxide RRAM element	rram	2T	resistor, crossbar	filamentary, valence-change, or unknown
PCM/PRAM cell	pcm, pram	2T	resistor	phase-change, thermal
ferroelectric capacitor	fram	2T	capacitor	ferroelectric
FeFET test structure	fefet	3T	fet, tft, transistor	ferroelectric
ECRAM transistor	ecram	3T	tft, transistor	electrochemical, ion-insertion, redox
generic memory TFT	memtransistor or charge-trap-memory	3T	tft, transistor	charge-trap, floating-gate, or unknown
dual-gate memtransistor	memtransistor	4T	tft, transistor	evidence-dependent

A 3-terminal oxide ECRAM, for example, may be represented as:

technology: ecram
platform:   tft, transistor, three-terminal
mechanism:  electrochemical, ion-insertion, redox
material:   oxide, electrolyte
behavior:   analog, multilevel, nonvolatile
experiment: pulse-update

The tag groups are:

• technology_tags: RRAM/ReRAM, PRAM/PCM, FRAM, FeFET, FTJ, ECRAM, CBRAM, MRAM, flash, memtransistor, and related concepts;
• terminal_count: integer number of electrically accessible DUT terminals;
• platform_tags: resistor, capacitor, diode, transistor, FET, TFT, MOSFET, OECT, crossbar, 1T1R/1T1C, and geometry tags;
• mechanism_tags: filamentary, valence-change, ECM, phase-change, ferroelectric, electrochemical, ionic, charge-trap, magnetic, tunneling, interfacial, and unknown/mixed;
• material_tags: oxide, chalcogenide, ferroelectric oxide, organic, 2D, electrolyte, silicon, and related systems;
• behavior_tags: analog, binary, multilevel, volatile/nonvolatile, threshold switching, stochastic, synaptic, and other observed behavior;
• experiment.profile: pulse update, DC I–V, transfer/output, retention, endurance, polarization, transient, impedance, noise, and related measurement types.

terminal_count refers to the interface of the measured DUT, not automatically to a circuit-cell label such as 1T1R or 1T1C. Circuit integration is recorded with separate platform tags.

Why this project exists

Device papers frequently use incompatible column names, terminal conventions, state definitions, pulse timing, normalization choices, and energy assumptions. The problem becomes particularly serious when a two-terminal programming path and a three-terminal gate/electrolyte path are analyzed with the same current model.

MemDeviceBench provides:

• a canonical long-form CSV schema for pulse-by-pulse read conductance;
• a JSON metadata sidecar with controlled multi-axis tags;
• illustrative device-profile templates for RRAM, PCM/PRAM, FRAM, FeFET, ECRAM, TFT, OECT, MRAM, FTJ, and 4T memtransistors;
• validation for units, branch order, signs, timestamps, terminal-tag consistency, and read-current/read-voltage consistency;
• transparent linearity, monotonicity, update variability, symmetry, conductance-window, endurance, retention, and energy metrics;
• topology-aware energy handling that prevents channel read conductance from being substituted for an unmeasured 3T/4T programming current;
• multi-terminal programming-path provenance warnings that compare generic pulse current with terminal-resolved and read-path current when available;
• machine-readable JSON/CSV reports and an overview figure;
• contribution units for instrument importers, metrics, public datasets, and future experiment profiles.

Quick start

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e .

Inspect the vocabulary and common tag templates:

memdevice-bench taxonomy
memdevice-bench device-profiles
memdevice-bench presets

Generate and analyze a 2-terminal tutorial trace:

memdevice-bench generate \
  --preset rram-2t \
  --output examples/rram_2t.csv \
  --cycles 4 \
  --pulses 64

memdevice-bench validate \
  examples/rram_2t.csv \
  --metadata examples/rram_2t.metadata.json

memdevice-bench analyze \
  examples/rram_2t.csv \
  --metadata examples/rram_2t.metadata.json \
  --output-dir examples/rram_2t_report

Create metadata from a profile template, then edit it for the actual sample:

memdevice-bench init-metadata \
  --template ecram-tft-3t \
  --device-id wafer03-d17 \
  --output wafer03-d17.metadata.json \
  --instrument "Keithley 4200A-SCS"

Or specify every axis manually:

memdevice-bench init-metadata \
  --output device.metadata.json \
  --device-id device-001 \
  --terminals 4 \
  --technology memtransistor \
  --platform tft transistor \
  --mechanism charge-trap \
  --material oxide two-dimensional \
  --behavior analog multilevel nonvolatile \
  --instrument "custom pulse/SMU setup"

Python API:

from memdevice_bench import (
    analyze_trace,
    generate_synthetic_trace,
    metadata_for_preset,
)

trace = generate_synthetic_trace(
    preset="rram-2t",
    cycles=3,
    pulses_per_branch=64,
    seed=7,
)
metadata = metadata_for_preset("rram-2t")
report = analyze_trace(trace, metadata=metadata)
print(report.summary["all_tags"])

Canonical pulse table

Required columns:

Column	Meaning
device_id	Stable DUT identifier
cycle	Non-negative update-cycle index
pulse_index	Pulse order within each branch
direction	potentiation or depression; common aliases are normalized
conductance_s	Positive read-conductance magnitude in siemens

Recommended generic columns:

Column	Meaning
pulse_voltage_v	Signed voltage across the programming path
pulse_width_s	Programming-pulse duration
pulse_current_a	Current in the programming path
read_voltage_v	Read-path voltage
read_current_a	Read-path current
timestamp_s	Elapsed experiment time
pulse_terminal	Programming path, e.g. gate-to-source
read_terminal	Read path, e.g. drain-to-source

Optional terminal-resolved columns include gate, drain, source, and body/back-gate voltages and currents. All canonical numerical fields use SI units.

Topology-aware programming energy

The preferred energy expression is measured-path integration:

E = |V_program I_program t_pulse|

A V_program² G_read t_pulse fallback is allowed only when metadata explicitly identifies a 2-terminal DUT. It is never silently applied to an unknown topology. For a known 3T/4T DUT, missing programming-path current produces energy_status = not_computed rather than substituting channel read conductance.

Even measured VIt remains an approximation when a scalar pulse current hides RC transients, displacement current, ionic relaxation, compliance clipping, waveform distortion, or finite-bandwidth sampling. Waveform-level integration is a roadmap item.

Implemented pulse-update metrics

• conductance window and dynamic-range ratio;
• endpoint-normalized linearity NRMSE and linear-fit R²;
• monotonic update fraction;
• mean/median update magnitude and update coefficient of variation;
• early-to-late update ratio as a saturation indicator;
• potentiation/depression update asymmetry;
• cycle-to-cycle conductance-window drift;
• programming energy with explicit provenance;
• power-law retention exponent and fit quality.

No universal device-quality score is produced. Cross-technology comparison must retain pulse amplitude, width, read bias, temperature, topology, state-preparation, and metric-definition context.

Synthetic presets

The package includes software-test traces for:

• rram-2t
• pcm-pram-2t
• cbram-2t
• ecram-tft-3t
• fefet-tft-3t
• memtransistor-4t

These traces test schemas and workflows. They are not compact physical models or literature benchmark data. FRAM is represented in the metadata templates, but no conductance-trajectory preset is presented as a substitute for polarization/PUND data.

Documentation and projectization

• Korean overview: README.ko.md
• tagging model: docs/tagging.md
• data format: docs/data-format.md
• metric definitions and limitations: docs/metrics.md
• architecture: docs/architecture.md
• Korean project brief: docs/project-brief-ko.md
• reusable Korean project-builder prompt: prompts/MEMDEVICEBENCH_PROJECT_PROMPT_KO.md
• public launch plan: PROJECT_LAUNCH.md
• community contribution map: COMMUNITY_TASKS.md

Development

pip install -e ".[dev]"
ruff check .
mypy src/memdevice_bench
pytest
python -m build
python -m twine check dist/*

Repository links target the JK1234567899999/memdevice-bench public project. Verify the package name on the target registry before publishing a package release.

License

Apache-2.0. Dataset licenses remain dataset-specific and must be declared in each metadata sidecar and dataset card.