Performance optimization in thermoelectric coolers using physical data-driven artificial neural networks

Driven by industrial growth and the demand for clean energy, thermoelectric coolers (TECs) have emerged as a sustainable, pollution-free cooling solution; however, their widespread adoption remains limited by low coefficient of performance (COP) under varying thermal and electrical conditions. This study presents a deep learning framework based on artificial neural networks (ANNs) developed in MATLAB to predict TEC performance across diverse operating scenarios. A dataset of 40,000 samples was generated from fundamental physical equations with optimized electrical current values obtained via an iterative gradient descent, ensuring adherence to the thermoelectric nonlinearities and energy conservation. The ANN architecture comprises one input layer, 15 hidden layers with Leaky Rectified Linear Unit (ReLU) activation, and one output layer, trained using the Root Mean Square Error (RMSE) as the loss function. To enhance predictive accuracy, a Genetic Algorithm (GA) was employed to optimize key geometric parameters-leg length, leg width, and fill factor. The optimized ANN-GA model achieved RMSE reduction of approximately 60 % under constant cold-side temperature (T_c) and 54 % under constant heat-flux (Qc˙) conditions, maintaining coefficient of determination (R²) values above 0.95 for both COP and power density (PD) predictions. The strong agreement between predicted and actual results confirms the model’s robustness and accuracy. This study establishes a physics-informed, data-driven ANN-GA framework for reliable performance prediction and design optimization of TECs, providing a foundation for future advances in thermoelectric device development and sustainable cooling technologies.