An Efficient Machine Learning Approach for Degradation Forecasting in AEM Water Electrolysis

I. INTRODUCTION

• Limited long-term membrane stability is a major barrier to large-scale AEMWE deployment, motivating predictive health and performance monitoring.The goal is to optimize maintenance and avoid critical failures.Source: §I · p. 1
• AEMWE machine-learning studies have mainly addressed performance prediction and operating-point optimization, with only limited work on forecasting degradation trajectories.Prior degradation modeling has also used physics-based approaches, which require material characterization and calibration and often assume slowly varying conditions.Source: §I · pp. 1–2
• The study introduces AEMWE degradation trajectories capturing transient early-life dynamics across heterogeneous experimental runs.The dataset is intended to support predictive models for degradation dynamics on a budget and insights applicable to industrial settings.Source: §I · pp. 1–2
• The proposed training and evaluation framework uses Leave-One-Group-Out cross-validation for data-scarce settings and industrial deployment.Source: §I · p. 2
• The study compares linear and nonlinear baselines, shallow regularized networks, LSTMs, and 1D-CNNs to assess expressive power versus stability.Source: §I · p. 2

II. EXPERIMENTAL DATASET AND PROBLEM DEFINITION

• The dataset contains 12 constant-current aging runs on a 5 cm^2 AEM single cell, each lasting 40–100 hours.The tests used heterogeneous operating protocols and focused on cell-voltage evolution.Source: A. Experimental Dataset and Notation · p. 2
• Figure 1 presents a representative raw voltage profile from one experiment.Source: Figure 1 · p. 2
• Each sample records cell voltage at a timestamp alongside static metadata describing initial experimental conditions.Examples of metadata include membrane properties and catalyst layers.Source: A. Experimental Dataset and Notation · p. 2
• The modeling pipeline omits metadata and forecasts voltage as a univariate time series because preliminary tests found no benefit or decreased performance from including those features.The raw voltage profiles show complex nonlinear dynamics with distinct local trends.Source: A. Experimental Dataset and Notation · p. 2

B. Preprocessing

• Irregular sampling and significant high-frequency noise motivate preprocessing each experiment into a uniform, noise-robust representation.Source: B. Preprocessing · p. 2
• Gaussian smoothing attenuates high-frequency noise while preserving long-term degradation trends.Source: B. Preprocessing · p. 2
• A cubic interpolant is fitted to the smoothed signal, then uniform resampling produces a discrete-time series.Source: B. Preprocessing · p. 3
• The fixed sampling interval is five minutes because degradation evolves on timescales significantly longer than one hour, making finer sampling unnecessary for this task.The number of samples may vary with experiment duration.Source: B. Preprocessing · p. 3

C. Final dataset construction

• Fixed-size input-output pairs are built from heterogeneous experiments of varying length using a sliding-window approach.Windows are highly overlapping, and all windows from an experiment stay in the same evaluation split to prevent data leakage.Source: C. Final dataset construction · p. 3
• The direct multi-horizon task predicts a scalar summary of the future state by averaging over the final portion of the prediction horizon.This averaging yields a robust estimate of the future regime and mitigates high-frequency noise.Source: C. Final dataset construction · p. 3
• Per-window standardization uses each input window’s mean and standard deviation to address experiment-level distribution differences and avoid data leakage.The local standardization emphasizes relative temporal dynamics and helps account for differences in absolute voltage scale.Source: C. Final dataset construction · p. 3
• The experiment index is retained in the final dataset, while each experiment can have a different range of window indices.The window length corresponds to a full 24-hour cycle, and the chosen stride creates highly overlapping windows.Source: C. Final dataset construction · p. 3

III. METHODOLOGY AND ARCHITECTURES

• The final dataset contains approximately 200–500 samples depending on the prediction horizon, creating a severe low-data regime.The setting carries high overfitting risk and makes careful control of model complexity essential.Source: §III · p. 3
• The study’s objective in this setting is to assess the robustness and relative inductive biases of modeling approaches, not to pursue state-of-the-art performance through extensive tuning.Source: §III · p. 3
• The study uses strict cross-experiment evaluation together with a restrained modeling strategy to address the low-data challenge.Source: §III · p. 3

A. Evaluation Protocol

• With 12 independent runs and highly overlapping samples, random splits could place samples from one experiment in both training and test sets, causing leakage and overly optimistic estimates.Source: A. Evaluation Protocol · p. 3
• Leave-One-Group-Out Cross-Validation assigns each experimental run to its own group and holds one run out as the test set in each fold.Source: A. Evaluation Protocol · p. 3
• Each fold trains from scratch on samples from the other experiments and evaluates on the held-out experiment; reported performance is averaged across folds to estimate cross-experiment generalization.Source: A. Evaluation Protocol · p. 3

B. Validation and Model Selection

• No additional validation split is used because limited experiments make estimates high-variance and model selection unstable across validation choices.Source: B. Validation and Model Selection · p. 3
• Table I summarizes model performance across time horizons using MAE.Source: Table I · p. 4
• Architectures and hyperparameters are fixed across folds, avoiding extensive tuning in favor of comparing model robustness and relative inductive biases in a low-data regime.Source: B. Validation and Model Selection · p. 4

C. Models and Architectures

• The model hierarchy ranges from linear baselines to deep neural architectures, trained with mean squared error; neural models use Adam, learning-rate scheduling, and early stopping.Source: C. Models and Architectures · p. 4

1) Baselines:

• The baselines include a global linear trend extrapolated from the input window, ridge regression with L2 regularization (α = 1), and a Random Forest regressor.The Random Forest uses 300 trees and a minimum leaf size of 5.Source: 1) Baselines · p. 4

• The shallow fully connected ReLU network has one 128-unit hidden layer, L2 normalization, and dropout at rate 0.2.Source: 2) Shallow neural network · p. 4

3) Deep temporal models:

• The 1D-CNN uses one 16-filter convolutional layer and max pooling, while the LSTM summarizes its input sequence through the final hidden state.The CNN then uses dropout and fully connected layers; the LSTM is intentionally small because of low data, and greater capacity degraded performance, suggesting overfitting sensitivity.Source: 3) Deep temporal models · p. 4

A. Performance metrics

• Performance is evaluated using Mean Absolute Error (MAE), comparing ground-truth targets with model predictions.Source: A. Performance metrics · p. 4
• MAE is reported in normalized units and centiVolts, with standard deviations across five training and evaluation runs.Physical-unit reporting supports interpretation in the context of electrochemical performance and degradation.Source: A. Performance metrics · pp. 4–5

B. Results

• Ridge Regression is the best-performing baseline by a clear margin; the global linear trend struggles with degradation complexity, and Random Forest weakens at longer horizons.Source: B. Results · p. 5
• The Shallow Neural Network is most consistent across time scales and has lower run-to-run standard deviation than temporal models; the 1D-CNN is stable but slightly less accurate.The LSTM performs well at longer horizons but degrades on shorter windows, and appears more sensitive to capacity and training choices.Source: B. Results · p. 5, §V · p. 5
• Centivolts MAE remains bounded across forecasting horizons without a systematic increase at longer timescales.The authors interpret this as stable accuracy in the original voltage scale.Source: B. Results · p. 5

V. CONCLUSIONS

• Simple regularized neural architectures offer the most reliable performance–stability trade-off under severe data scarcity and noisy electrochemical time series.The framework is described as robust to heterogeneous, data-scarce conditions and applicable to industrial settings.Source: §V · p. 5