Experimental Methodology

Overview

KTBIX employs a rigorous three-tier validation framework:

1. In-lab controlled experiments — Oxford & NASA datasets
2. Field validation — Turkish T-EV platform data (WP1)
3. Cross-chemistry generalization — Testing on unseen battery chemistries

EIS Data Acquisition Protocol

Equipment

• EIS analyzers: Gamry Reference 3000 / Zahner Zennium (ITU laboratory)
• Frequency range: 10 mHz – 100 kHz (51 frequency points per spectrum)
• Voltage amplitude: 10 mV RMS (quasi-linear regime)
• Temperature control: 15°C, 25°C, 35°C, 45°C (characterization) + ambient field conditions

Battery Platforms

Platform	Chemistry	Nominal Capacity	Source
T-EV Fleet	NMC	50–100 Ah	Turkish EV operators
TOGG T10X	NMC	~75 kWh pack	TOGG (MOU pending)
Oxford Dataset	LCO	740 mAh	Oxford University
NASA PCoE	18650	2.0 Ah	NASA Prognostics

SOH Ground Truth

SOH is defined as capacity ratio: $\text{SOH} = C_{\text{current}} / C_{\text{nominal}} \times 100\%$

Measured via CC-CV charge/discharge cycling with precision coulometry (±0.02% accuracy).

Comparison Baselines

Per the evaluation criteria, KTBIX validates against ≥5 state-of-the-art methods:

Method	Reference	Category
SVR + EIS features	Roman et al. (2021)	Classical ML
LSTM-SOH	You et al. (2017)	Deep learning
Transformer-EIS	Tian et al. (2020)	Attention-based
AE-BPNN	Güneşer et al. (2025)	Baseline
EIS EC-fitting (Randles)	Vetter et al. (2005)	Physics-based
SHAP-SVR	Hannan et al. (2021)	XAI-enhanced ML

Explainability (XAI) Framework

SHAP Analysis

SHAP (SHapley Additive exPlanations) attributions identify which impedance frequencies contribute most to SOH predictions:

import shap

explainer = shap.DeepExplainer(pi_vib_resnet_model, background_data)
shap_values = explainer.shap_values(test_eis_spectra)
shap.summary_plot(shap_values, feature_names=frequency_labels)

Gradient Saliency

Gradient-based saliency maps highlight input-space sensitivity regions corresponding to $R_{ct}$, Warburg element (W), and double-layer capacitance (CPE).

Evaluation Metrics

$\text{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}(\hat{y}_i - y_i)^2}$

$\text{MAE} = \frac{1}{N}\sum_{i=1}^{N}|\hat{y}_i - y_i|$

$R^2 = 1 - \frac{\sum(\hat{y}_i - y_i)^2}{\sum(y_i - \bar{y})^2}$

Target: RMSE ≤1.2%, MAE ≤0.9%, R² ≥0.98.

Risk Management

Risk	Probability	Mitigation
TOGG data access delay	Medium	Fallback: expand T-EV fleet + Zenodo public datasets
Model overfitting	Low	VIB compression + dropout + cross-validation
Cross-chemistry failure	Medium	Physics loss term constrains domain shift
Compute cost	Low	SJU HPC cluster + AWS spot instances