MagBridge-Battery: A Synthetic Bridge Dataset for Li-ion Magnetometry and State-of-Health Diagnostics | Sakthi Prabhu Gunasekar, Prasanna Kumar Rangarajan | ResearchPod

Transcript: MagBridge-Battery: A Synthetic Bridge Dataset for Li-ion Magnetometry and State-of-Health Diagnostics

Alex: Welcome to another episode of ResearchPod. Today we're discussing the MagBridge-Battery dataset — a new resource for developing magnetic-sensing diagnostics for lithium-ion batteries.

Sam: So the core problem is that magnetometry can reveal internal degradation that terminal sensors miss entirely, but there's no public dataset pairing those magnetic scans with state-of-health labels?

Alex: That's exactly it. And without that pairing, you can't train or benchmark any diagnostic model. You're stuck. The field has the sensing modality but no common ground truth to work against.

Sam: So how does the "bridge" actually close that gap?

Alex: The mechanism they call "retrieve-and-blend" starts from real magnetic morphology pulled from a public archive. It then morphs those base scans by blending in features weighted according to a battery's state-of-health label. So you're not generating from scratch — you're anchoring to real physical measurements and perturbing them in a label-informed direction.

Sam: That's the part I'd want to stress-test. How do you distinguish genuine label-aligned physical signal from the model just hallucinating artifacts that happen to correlate with the label?

Alex: That's the load-bearing question, and they address it directly with a label-shuffle ablation. When you randomize the SOH labels, model performance collapses. That's your evidence that the bridge is encoding a non-trivial relationship between the magnetic morphology and the health state — not just fitting noise.

Sam: So the generation is constrained to a subspace derived from real data, not free to roam?

Alex: Right. They project the base scans into a low-dimensional LDA subspace — derived from real magnetic anchors — and perturb along a fitted state direction within that subspace. The key word is constrained. Synthetic outputs are forced to respect the physical manifold of the source data. That's what separates this from unconstrained generative augmentation.

Sam: And the leakage-safe protocol — zero cell overlap between train and test splits — that's what makes the benchmark numbers actually trustworthy?

Alex: Exactly. It's a detail that matters enormously. If cells appear in both splits, you're measuring memorization, not generalization. The fact that they explicitly enforce and quantify that separation gives the benchmark real validity.

Sam: So what's the honest scope of the contribution here?

Alex: It's a synthetic bridge, not a substitute for future paired experimental data — and the authors are clear about that. What it does provide is a rigorous baseline where none existed before. Researchers can now compare diagnostic architectures on a common, standardized footing, and ask a question that previously had no controlled setting to answer: does magnetic sensing actually outperform terminal-only voltage measurements? That's a meaningful step for a nascent field. Thanks for listening to ResearchPod.

MagBridge-Battery: A Synthetic Bridge Dataset for Li-ion Magnetometry and State-of-Health Diagnostics

Sakthi Prabhu Gunasekar, Prasanna Kumar Rangarajan

Sep 3, 2026·2:42·user_upload

Battery State-of-Health (SOH) DiagnosticsMagnetometrySynthetic DatasetLeakage-safe SplitMagnetic-Electrochemical BridgingState of Health (SOH) Estimation

Listen

3 min

Briefing

Research Context

Battery health diagnostics are currently dominated by electrochemical measurements (voltage, current, impedance) taken at cell terminals. While these methods are well-established, they are inherently blind to internal spatial inhomogeneities like localized hotspots, uneven redox reactions, or internal defects. A growing body of research suggests that magnetic sensing can resolve these missing details, but progress has been stalled by the total absence of public datasets that pair magnetic measurements with degradation labels.

Approach

To bridge this gap, the authors developed MagBridge-Battery v1.0. This synthetic dataset combines real magnetic morphology from the Mohammadi–Jerschow Open Science Framework (OSF) archive with real state-of-health (SOH) labels from the PulseBat dataset. The bridge uses a deterministic, reproducible architecture consisting of a regime classifier, a morphology bank, a degradation modulator (powered by a quantum reservoir computer), and a noise model. The release includes 6,760 samples, categorized into grounded samples, synthetic sensor anomalies, and low-voltage extrapolation samples.

Main Findings

The dataset provides a rigorous benchmark for three primary tasks: SOH regression, second-life classification, and anomaly detection. To ensure the dataset is not merely producing label-aligned artifacts, the authors performed a series of controlled ablations. Most notably, a label-shuffle test—where SOH labels are permuted before generation—caused SOH regression performance to collapse from an R-squared of approximately 0.77 to near zero. This confirms that the bridge successfully encodes meaningful, SOH-dependent information into the synthetic magnetic signatures.

Why It Matters

MagBridge-Battery v1.0 provides the first public benchmark for researchers working on magnetic-sensing battery diagnostics. By offering a standardized, leakage-safe protocol (using a cell-disjoint split), it enables cross-lab comparison and method development without requiring access to proprietary or scarce paired magnetic-electrochemical data. It serves as a crucial placeholder until large-scale, real-world paired datasets become available.

Key Terms and Definitions

MagBridge-Battery v1.0 — A synthetic dataset of 6,760 magnetic-field signatures paired with electrochemical degradation labels designed for battery health diagnostics.
Grounded Regime — The primary operating voltage range (3.06 V to 3.34 V) where both magnetic morphology and electrochemical SOH labels are available for conditioning.
Regime-B — A low-voltage extrapolation range (2.54 V to 3.00 V) included for out-of-distribution and anomaly detection evaluation, where SOH labels are intentionally omitted.
MagBridge-Embed — A 171-dimensional latent representation produced by a quantum reservoir computer used to modulate degradation signals.
Cell-Disjoint Split — A data partitioning strategy that ensures no physical battery cell appears in more than one subset (train, validation, or test), preventing data leakage.

MagBridge-Battery: A Synthetic Bridge Dataset for Li-ion Magnetometry and State-of-Health Diagnostics: Research Paper Podcast