Innovative approaches for predicting cryogenic hydrogen behaviour

Safe and efficient methods for storing and transporting liquid hydrogen are crucial for decarbonisation. This study presents an advanced method for analysing cryogenic hydrogen releases, combining large eddy simulation (LES), proper orthogonal decomposition (POD), and machine learning (ML) models to predict hydrogen dispersion and temperature distribution. Using LES data to model Sandia’s cryogenic hydrogen jet experiment (5 bar, 50 K), 540 snapshots of hydrogen mole fraction and temperature fields are extracted over a time range of 0.15 s to 0.42 s. POD identifies the first 10 dominant spatial modes, reducing the dataset dimensionality while preserving critical flow structures. A bidirectional long short-term memory (BiLSTM) model is adopted to predict the subsequent temporal coefficients with high accuracy for both hydrogen mole fraction and temperature distribution. For the mole fraction, the root mean squared error (RMSE) ranges from 0.017 to 0.059, the coefficient of determination (R2) varies between 0.944 and 0.997, and the correlation coefficient (R) remains between 0.972 and 0.999. Similarly, for temperature, RMSE spans from 0.024 to 0.061, R2 ranges from 0.941 to 0.993, and R varies between 0.971 and 0.998. The reconstructed results closely match LES data, with the ML-based method achieving a lower deviation (2.85 %) in predicting the hydrogen flammability threshold (0.04 mol fraction) compared to POD (3.56 %). The model accurately identifies the extent of the flammable region, which is essential for fire risk assessment and determining safe separation distances. In addition, this method identifies regions with increased ignition potential due to high hydrogen concentration and temperature gradients, helping in hazard mitigation strategies. The research contributes to safe hydrogen storage and transport, supporting global decarbonisation efforts.