Optimisation of pin-fin heat sink design for CPV systems using machine learning-driven multi-objective approaches

Concentrated photovoltaic (CPV) systems, with their high efficiency and compact design, support green hydrogen production and contribute to United NationsSustainable Development Goal 7 (Affordable and Clean Energy). However, their performance and longevity can be significantly compromised by inadequate thermal management. To address this challenge, this study proposes a data-driven framework that enhances CPV thermal optimisation while reducing reliance on computationally intensive simulations. A novel variable pin–fin height heat sink is evaluated with the aim of minimising maximum temperature rise, temperature non-uniformity, and pressure drop. Five tree-based machine learning (ML) models, including Decision Tree, Random Forest, Gradient Boosting, XGBoost, and CatBoost, are assessed, with CatBoost demonstrating the highest predictive accuracy. This model is used as a surrogate for high-fidelity simulations, enabling efficient multi-objective optimisation using the Non-dominated Sorting Genetic Algorithm III (NSGA-III). To identify the most balanced configuration among the Pareto-optimal solutions, six multi-criteria decision-making (MCDM) models are applied. Results indicate that PROBID and PROMETHEE II, among the six MCDM models, effectively balance the competing objectives, producing a configuration that reduces temperature non-uniformity by 70.48 % and pressure drop by 41.84 % compared to a conventional uniform design. Validation against high-fidelity simulations confirms the accuracy of ML predictions, with an error margin between 0 and 0.063 %. This integrated surrogate-based approach offers a cost- and time-efficient solution for optimising CPV systems and other high-density thermal applications.