Numerical ISPH–AI study of heat transfer and entropy generation in a porous C-shaped enclosure filled with Casson–NEPCM slurry

Compact electronic packages and miniaturized thermal devices increasingly operate under intermittent heat release, where a coolant must absorb local heat pulses while preserving a small temperature margin. Porous inserts and NEPCM slurries are promising for this purpose, yet their combined behavior is difficult to assess when buoyancy, permeability drag, Casson rheology, latent heat, nonlinear radiation, and second-law losses act at the same time. The C-shaped enclosure is selected as an idealized compact-cooling benchmark because its re-entrant geometry, confined passages, and internal finned heat source create nonuniform thermal paths similar to those encountered in miniaturized electronic packages. Based on the literature reviewed here, these mechanisms are addressed together in one laminar natural-convection benchmark, without imposing any inlet, outlet, moving wall, or forced-flow condition. The governing equations combine Darcy–Brinkman–Forchheimer momentum resistance, an effective Casson viscosity, an apparent-heat-capacity NEPCM model, viscous dissipation, nonlinear Rosseland radiative diffusion without T 4 linearization, and entropy-generation decomposition. Transient fields are computed with an incompressible smoothed particle hydrodynamics formulation over R a = 1 0 3 –10 6 , D a = 1 0 − 5 – 1 0 − 2 , β c = 0.01 –10, R d = 0 –4, E c = 0 –0.10, ϕ = 0 –0.06, and θ f = 0.05 –0.9. The resulting ISPH database is then emulated with stacked machine-learning surrogates formed from Extra Trees and histogram-gradient boosting regressors with Ridge regression as the meta-learner. These AI predictions are interpolation tools trained on ISPH-generated data and do not replace the physical simulations, governing equations, or validation tests. The boundary-condition comparison gives the largest thermal contrast: Outer Cold/CPU Hot produces the strongest N u ¯ response, Outer Hot/CPU Cold is intermediate, and Center Hot/CPU Cold is the weakest arrangement. Increasing D a from 1 0 − 5 to 1 0 − 2 relaxes porous resistance and extends the active transport region, whereas increasing R d from 0 to 4 monotonically strengthens N u ¯ through radiative heat spreading. Entropy generation rises in configurations with steeper imposed thermal gradients and at stronger radiative participation, while the Bejan-number maps show that heat-transfer irreversibility controls most of the cavity except near zones where porous drag and velocity gradients intensify. For N u ¯ prediction, the stacked model reaches R 2 = 0.999995 and RMSE = 1.571 × 1 0 − 2 in the boundary-condition dataset, and R 2 = 0.999996 and RMSE = 3.340 × 1 0 − 2 in the radiation dataset. The results provide a physics-resolved and surrogate-assisted screening route for porous latent-heat cooling concepts. They indicate how permeability, radiation level, thermal-boundary placement, and NEPCM activation can be coordinated before moving to costlier three-dimensional simulations, experiments, or device-specific optimization.