Data-driven forecasting of solid oxide electrolysis cell degradation using deep transfer learning

Solid oxide electrolysis cells (SOECs) offer high efficiency for hydrogen production but can suffer from progressive performance degradation that limits durability and commercial deployment. Reliable forecasting of long-term degradation is therefore critical for condition monitoring, lifetime management, rapid material/cell development, and operational decision-making. However, existing physics-based models require extensive parameterization, while most data-driven approaches focus on short-term performance prediction and struggle to generalize across cells, particularly when degradation data are limited.In this work, a data-driven prognostics framework was developed to forecast SOEC degradation trajectories using a hybrid convolutional neural network-long short-term memory (CNN-LSTM) architecture augmented with an attention mechanism and deep transfer learning. The model was trained on degradation data from 8 source cells and evaluated on 12 independent cells tested under varying operating conditions and designed to capture both local temporal patterns and long-range dependencies in bounded degradation signals. Transfer learning was employed to adapt the pretrained model to previously unseen cells using only a small fraction of early-life data. Cross-validation on the source dataset demonstrated high predictive accuracy and strong agreement between predicted and measured degradation trajectories. When applied to independent target cells, the transfer learning strategy reduced RMSE from 0.0588 to 0.0130 on unseen cells when only 10% of early-life data was available. These results demonstrate that deep transfer learning enables accurate early-life, long-horizon degradation forecasting for SOECs, offering a practical pathway toward data-efficient durability prediction and improved operational management of electrolysis systems.