Failure prediction in composite hydrogen storage tanks using recurrent deep learning models

A mechanics-based data-driven deep learning framework is proposed to evaluate the mechanical response of filament-wound composite hydrogen storage tanks under internal pressurization. Using Classical Laminate Theory (CLT), a synthetic, layer-resolved dataset is generated to capture stresses and strains as functions of stacking sequence and fiber orientation. Structural integrity is assessed via the Tsai–Wu failure criterion to identify configurations that remain within allowable limits. Several sequence-based models, including Long Short-Term Memory (LSTM) networks, are benchmarked to predict the layer-wise Tsai–Wu index under prescribed pressure ramps. The optimal LSTM model achieves an R 2 of 0.95 and an RMSE of 0.018 on the test set. The predictions remain consistent across layers and pressure levels, and comparisons with numerical simulations from WoundSim and Abaqus show close agreement, confirming the robustness of the approach. By integrating mechanics-based data generation with deep learning, the proposed framework accelerates the design and optimization of composite pressure vessels while significantly reducing computational and training costs for industrial applications.