Zenthos Energy Battery Research Modeling and Historical Data Processing initiative
Document
Classification1. Project Scope The objective of this project is to build an end-to-end data pipeline that digitizes, curates, and labels historical battery testing datasets to train predictive machine learning models. Focusing heavily on advanced, non-lithium architectures—specifically aluminum-air battery systems—the project transforms decades of unstructured hardware testing data, legacy lab notes, and sensor logs into structured training sets. The downstream goal is to deploy Machine Learning (ML) models for State of Health (SoH) prediction, degradation modeling, and real-time optimization of electrolyte/air-cathode dynamics under varying industrial loads. 2. Project Size (Scale) Data Sources: Ingestion of continuous galvanic cycling time-series logs, Electrochemical Impedance Spectroscopy (EIS) sweep profiles, and historical thermal runaway test records. Volume: Aggregating and structuring over 1.5 million hours of cumulative cell-level and pack-level historical cycling data spanning different anode compositions, mechanical air-cathode configurations, and carbon dioxide management profiles. Feature Density: Multi-variate time-series streams sampling voltage, current, internal resistance, structural expansion, and ambient/localized temperature at frequencies up to 10 Hz. 3. Specific Data Labeling Tasks Performed To transition raw electrochemical telemetry into a high-fidelity ML dataset, the annotation team executes several precise labeling primitives across different modalities: A. Time-Series Windowing & Event Labeling Classification (Anomalous Events): Annotators identify and explicitly label discrete electrochemical anomalies within continuous time-series logs, tagging events such as [Anode_Passivation], [Electrolyte_Depletion], [Carbonation_Choking], and [Micro_Short_Circuit]. Segmentation (Cycle-Phase Parsing): Programmatic and manual verification of exact cycle boundaries, separating dynamic discharge intervals, resting states, and mechanical electrolyte replenishment windows. B. Analytical Imaging & Spectroscopy Parsing Bounding Box / Polygon Annotation: Labeling scanning electron microscopy (SEM) and post-mortem physical images of spent aluminum anodes to categorize spatial degradation metrics (e.g., pitting depth, localized corrosion, and crystalline byproduct accumulation). Point / Curve Fine-Tuning (EIS Analytics): Highlighting and labeling characteristic inflection points, high-frequency arcs, and low-frequency diffusion tails on Nyquist plots generated during EIS testing to calibrate equivalent circuit models. C. Legacy Document NLP Layer Transcription & NER: Standardizing legacy, unstructured physical lab logs, slurry mixing formulations, and handwritten test parameters into uniform digital text formats, applying Named Entity Recognition (NER) to isolate [Anode_Purity_Grade], [Additive_Concentration], and [Catalyst_Load]. 4. Quality Measures Adhered To Because battery modeling demands absolute fidelity to prevent hazardous hardware failures and misleading lifecycle predictions, the data annotation pipeline maintains rigorous data hygiene standards: Electrochemical Sanity Constraints: The labeling platform enforces hard physical boundary checks based on first-principles battery physics (e.g., automatically rejecting or routing any transcribed open-circuit voltage data that violates thermodynamic limits or exhibits non-physical negative resistance). Multi-Expert Consensus Matching: Complex degradation categorization and structural SEM imaging analysis are subject to a blind dual-pass review process. Discrepancies between annotations are automatically escalated to senior battery engineers for definitive arbitration. Sensor Drift & Noise Filtering: All raw time-series inputs undergo automated baseline corrections and Kalman filtering to isolate true cell performance from external sensor noise, temperature swings, or wire resistance discrepancies prior to final human validation.