Artificial Neural Network Surrogate for Entropy Generation Analysis in PCM-Based Thermal Energy Storage

Latent heat thermal energy storage (LHTES) systems based on phase change materials (PCMs) are central to improving energy efficiency in solar–thermal, waste-heat recovery, and building energy management applications; however, minimizing the thermodynamic losses that degrade their exergetic performance requires fast and accurate prediction of entropy generation across wide operating envelopes. This study develops a coupled numerical–surrogate framework for predicting entropy generation during one-dimensional conduction-dominated melting of a PCM slab. An explicit finite-difference enthalpy solver is formulated in fully dimensionless variables governed by five independent parameters: the Stefan number (Ste), Biot number (Bi), initial subcooling (θi), dimensionless temperature difference (Π), and solid-to-liquid conductivity ratio (rk). The solver is verified against the analytical Neumann solution, achieving mean front-position errors below 0.7%. A Latin Hypercube Sampling campaign of 400 parameter combinations yields 378 converged cases spanning nearly three orders of magnitude in the entropy generation number Ns. A compact multilayer perceptron (5→64→64→32→1; 6657 parameters) trained on these data reproduces Ns with a test-set coefficient of determination R2=0.991 and a mean absolute percentage error of 5.73% while delivering a speed-up of approximately four orders of magnitude over the full solver. A global sensitivity analysis reveals that the dimensionless temperature difference Π is the dominant driver of entropy generation (Spearman ρ=+0.80), followed by the Stefan number (ρ=−0.51) and the Biot number (ρ=+0.25), whereas initial subcooling and conductivity ratio exert only marginal influence. The framework enables rapid, second-law-aware screening of PCM operating conditions, supporting energy-efficient design and optimization of thermal energy storage systems.