Physics-informed latent-space neural surrogate for fast liquid-hydrogen heat transfer prediction in hydrogen-powered cryo-electric aviation

The aviation industry's shift toward zero-emission hydrogen propulsion relies critically on the precise thermal management of cryogenic liquid hydrogen (LH 2 ) to effectively monitor boil-off and two-phase flow. This paper proposes a deployment-oriented, lightweight intelligent predictive framework that combines a Cascade-Forward Neural Network (CFNN) with physics-informed feature extraction and dimensionality reduction. By compiling a comprehensive steady-state experimental measurement database consisting of 3131 points, spanning different fluid regimes from subcooled to superheated fluid, 10 fundamental dimensionless parameters were extracted to capture the complex multi-phase fluid dynamics. The framework systematically benchmarks Principal Component Analysis (PCA), Partial Least Squares (PLS), and Autoencoders to compress this high-dimensional input space. Results demonstrate that PLS provides a highly efficient latent structure, achieving stable predictive accuracy with a compact number of latent variables. The integration of these linear dimensionality reduction techniques successfully lowered the network's computational parameters, reducing them from 4100 to 3209 in the most complex scenario. Consequently, the proposed hybrid models yielded inference speed-ups of up to 11.236% compared to baseline networks, reducing execution time to approximately 4.3 milliseconds. This computationally efficient framework preserves the high fidelity of deep learning surrogates, paving the way for real-time onboard monitoring and dynamic digital twins to optimise cryogenic thermal management in hydrogen-powered cryo-electric aircraft.