Physics-embedded neural modelling of thermo-mechanical fatigue in fibre composites: Interpretable parameters and uncertainty-quantified life

Predicting fatigue life of polymer-matrix composites with physically interpretable parameters and quantified uncertainty remains difficult on small datasets, where neither pure-physics nor pure-data-driven models suffice. This work presents a hybrid neural-physics framework that embeds a thermo-mechanical damage ordinary differential equation (ODE) in the forward path of a neural model. The ODE supplies physically meaningful parameters while the neural components capture residual variability, so that life predictions, thermo-mechanical trajectories, and material parameters are recovered jointly. A sliding-window long short-term memory (LSTM) encoder summarises cycle-resolved measurements, and a conditional variational latent represents the trajectory. A multi-channel decoder reconstructs the cyclic mechanical and energy signals, while a life-prediction head consumes the ODE end-state and is calibrated by split-conformal prediction. Material-intrinsic ODE parameters are shared across loading configurations and thermal-coupling parameters are configuration-specific. The framework is evaluated on 32 [±45]2s E-glass/epoxy specimens tested under load control at R = 0.1 in three configurations spanning two frequencies and two cooling conditions. For this system, the model predicts log10 Nf with a root-mean-square error of 0.49, and the 90 % split-conformal interval spans a factor of 4.8 in cycles with an empirical in-regime coverage of 0.84. Shared material parameters converge across 25 seed-and-fold runs with coefficients of variation at or below 1.4%. The framework matches the Basquin baseline on the headline life-prediction metric while additionally recovering interpretable material parameters and the temperature, modulus, and dissipated-energy trajectories. Coverage degrades outside the training regime and requires recalibration on new configurations.