IDEALLab/power_electronics_v1

EngiBench PowerElectronics v1

PowerElectronics v1 is a simulator-frozen EngiBench dataset for data-driven engineering design. It re-simulates the exact 13,824 input designs and preserves the train/validation/test membership from `IDEALLab/power_electronics_v0`, but it is a new benchmark and does not claim to reproduce the v0 labels.

The canonical backend is ngspice 44.2 on Linux x86_64 in a pinned Apptainer image. v1 corrects the objective semantics, retains signed physical measurements, and records the source, code, netlist, simulator, and container fingerprints in every row.

Loading

Pin the release tag in reproducible experiments:

from datasets import load_dataset

dataset = load_dataset(
    "IDEALLab/power_electronics_v1",
    revision="v1.0.0",
)

Splits and simulation status

The three invalid simulator results are retained at train[4309], train[5960], and val[223]. Their unavailable measurements and objectives are null; they are never replaced with zero, imputed, clipped, or removed.

Consumers should retain every row and use objectives_valid as a mask when a numeric regression loss requires objective labels. Invalid rows may also be used to model the simulator-failure region. Evaluation must report the number of valid and invalid rows rather than silently dropping failures.

All finite values, including negative signed measurements and very large relative-ripple values near zero mean voltage, remain unchanged.

v1 objectives

The stored physical measurements are:

• —signed mean output voltage, output_voltage_mean;
• —peak-to-peak output voltage, output_voltage_peak_to_peak.

With a 1,000 V source, v1 derives the signed gain and two minimization objectives:

dc_gain = output_voltage_mean / 1000
dc_gain_error = abs(dc_gain - 0.25)
relative_voltage_ripple = output_voltage_peak_to_peak / abs(output_voltage_mean)

If the simulator does not return finite measurements, the derived objectives are null and the validity/status fields identify the failure. A zero mean voltage would similarly make relative ripple unavailable rather than assigning an artificial value.

Fields

Identity and source mapping

• —dataset_schema_version: schema version for this row format.
• —split, source_index: exact v0 split membership and row index.
• —design_sha256: stable checksum of initial_design.
• —initial_design: the original 20-dimensional v0 design.
• —v0_DcGain, v0_Voltage_Ripple: immutable v0 labels retained only for direct comparison.
• —source_dataset_id, source_dataset_revision: exact v0 source.
• —problem_version: EngiBench problem version (1).

Reproducibility

• —engibench_git_commit, netlist_sha256;
• —simulator_version, simulator_sha256;
• —simulator_platform_system, simulator_platform_machine;
• —container_sha256.

Measurements, objectives, and validity

• —output_voltage_mean, output_voltage_peak_to_peak, dc_gain;
• —dc_gain_error, relative_voltage_ripple;
• —simulation_valid, objectives_valid, simulation_status;
• —error_type, error_message for caught simulator exceptions.

Frozen provenance

The exact original v0 runtime cannot be confirmed. An archived notebook contains evidence for Windows/x8664 and ngspice 36, while an ngspice 44.2 ARM64 Mac reproduction is much closer to the published v0 labels than the corresponding Linux x8664 reproduction. Because operating system, simulator version, compiler, and build also differ, the latter is evidence consistent with ARM but is not proof of the original architecture.

Do not combine v0 and v1 labels or silently train on one backend and verify on another. Cross-backend comparisons should be explicit experiments.

Generation and audit

The provenance/ directory contains:

• —the 24 original generation manifests;
• —the strict 13,824-row audit report;
• —the JSONL-to-Parquet semantic conversion report;
• —the container recipe and exact Python dependency freeze;
• —the Euler Slurm generation script and both audit/conversion scripts.

The Parquet conversion sorts every split by source_index. Its report verifies that every value and null round-trips exactly from the audited JSONL source.

Limitations

• —Results are specific to the frozen ngspice binary and container fingerprint.
• —Relative ripple can become extremely large when the absolute mean voltage is close to zero. These finite values are intentional and are not clipped.
• —Three designs do not have finite simulator measurements on the canonical backend. Their explicit invalid status is part of the benchmark.
• —The v0 labels are provided for comparison, not as additional v1 targets.

Citation

@inproceedings{felten_engibench_2025,
  title = {{EngiBench}: {A} {Framework} for {Data}-{Driven} {Engineering} {Design} {Research}},
  author = {Felten, Florian and Apaza, Gabriel and B\"{a}unlich, Gerhard and Diniz, Cashen and Dong, Xuliang and Drake, Arthur and Habibi, Milad and Hoffman, Nathaniel J. and Keeler, Matthew and Massoudi, Soheyl and VanGessel, Francis G. and Fuge, Mark},
  booktitle = {Proceedings of the 39th Conference on Neural Information Processing Systems ({NeurIPS} 2025)},
  year = {2025}
}