Accurate and efficient design and optimization of thermoelectric cooler via machine learning technologies

The increasing demand for efficient thermal management in applications such as on-chip hotspot cooling and battery systems calls for high-performance thermoelectric coolers (TECs) that can be rapidly optimized for varying thermal loads. In this work, we propose a data-driven framework that integrates a high-fidelity artificial neural network (ANN) with a genetic algorithm (GA) to enable fast and accurate prediction and optimization of TEC performance across a broad design and operating range. This approach is particularly valuable in real-world scenarios where input power fluctuates over time and maximizing TEC efficiency is critical. Unlike previous studies, which primarily relied on ANN models for performance prediction alone, the present work extends their functionality to include efficient configuration optimization within significantly shorter timescales. The ANN was trained on 20,000 data points generated from validated COMSOL Multiphysics® simulations under diverse conditions, achieving a prediction accuracy exceeding 99.9 %. Once trained, the ANN performs performance predictions over 180,000 × faster than COMSOL, enabling high-throughput simulation, sensitivity analysis, and real-time optimization. Coupling the ANN with a GA, we optimized TEC geometry and operating current to achieve minimum cold-side temperatures under varied boundary conditions. The ANN–GA framework shows excellent agreement with COMSOL-based optimization while achieving orders-of-magnitude gains in computational efficiency. Practical scenarios with dynamically varying heat inputs and temperature thresholds were also explored in which the system successfully adapts to changing loads, identifying optimal currents in under 33 s per case, demonstrating strong potential for real-time adaptive thermal management.