Nouha Arfaoui, R. Thaljaoui, Abdullah Aljaafari, Mohammad Sajid Mohammadi, Medhat Asem, Bedir Yousif
The excellent efficiency and environmental friendliness of the magnetocaloric effect (MCE) make it a promising alternative to conventional gas-compression refrigeration. However, the intricate relationship between structure, composition, and magnetic response complicates the identification and optimization of high-performance magnetocaloric materials. In this work, a machine-learning framework is developed to predict the maximum magnetic entropy change (ΔSₘₐₓ) of perovskite manganite compounds using only crystallographic lattice parameters (a, b, c) as input descriptors. The study relies on a curated dataset of 114 experimentally reported samples with corresponding magnetocaloric measurements. Model performance was assessed using MAE, RMSE, R², and the correlation coefficient (CC) across multiple regression algorithms. The results clearly show that Linear Regression and Ridge Regression provide the most accurate and stable predictions, achieving extremely low errors (RMSE ≈ 0.012 J·kg⁻¹·K⁻¹ and MAE ≈ 0.005) and near-perfect goodness-of-fit (R² and CC ≈ 0.9999). Their performance remains nearly identical with and without hyperparameter tuning, confirming strong robustness and indicating that the relationship between lattice parameters and ΔSₘₐₓ is predominantly linear. Other algorithms display contrasting behavior. KNN shows moderate performance with slight improvement after tuning, while the Gaussian Process model performs poorly without optimization but improves significantly once tuned. In contrast, K-Means and Naive Bayes exhibit weak regression capability, and the Dummy Mean model serves as a baseline reference. Overall, the findings demonstrate that simple linear models, when combined with physically meaningful structural descriptors, can accurately capture the structure–property relationship governing ΔSₘₐₓ, offering an efficient data-driven tool for magnetocaloric material screening.