In this contribution, a case study considering an unexpected corrosion-fatigue crack propagation issue in an aircraft fleet is used to discuss how to compensate for incomplete knowledge in time dependent responses integration and extrapolation. For the considered application, degradation resulting from mechanical fatigue is well understood and accounted in the damage models. However, the unexpected corrosion effects are not accounted in damage integration, yielding a large discrepancy between predicted and observed crack lengths. To address this epistemic uncertainty in the fleet damage accumulation model, hybrid neural networks cells are formulated; where physics-informed layers address well-understood aspects of the degradation, and data-driven layers are trained to act as correction terms. The considered case study encompasses highly imbalanced data sets with uncertainties acting asynchronously. To improve overall accuracy, ensemble learning techniques are adapted to merge the resulting hybrid neural network cells predictions. Lastly, a heuristic based on optimal ensemble weights is presented to help in the decision-making task of defining safe operation of the fleet. Results show that our proposed approach was capable of compensating for the epistemic uncertainties, and that the proposed heuristic can be used to rank aircraft damage severity, allowing to prioritize aircraft for inspection and/or route reassignment.