Application of Machine Learning for Estimating Heat Transfer Coefficients

The knowledge of the heat transfer coefficient plays a crucial role in evaluating coolants utilized for immersion quenching of steels. This coefficient effectively characterizes the heat exchange occurring between the immersed workpiece and the liquid coolant. The calculation of the heat transfer coefficient involves solving an inverse heat transfer problem, typically addressed using stochastic optimization algorithms. These algorithms rely on iterative processes and are computationally intensive, often requiring hundreds or even thousands of iterations to obtain a solution. To alleviate the computational burden, this paper introduces an initialization technique based on a non-iterative approach for solving the inverse heat transfer problem. The proposed method utilizes an artificial neural network to solve the problem. Specifically, a multi-layer feed neural network is utilized, trained using the backpropagation algorithm. In order to train the network, a synthetic database containing 150,000 records of heat transfer coefficients is created. The coefficient is determined as a function of temperature, with an unconventional utilization of the Fourier transform of the cooling curve as input for the inference system. Furthermore, the performance of the neural network is compared with other conventional learning algorithms. It is observed that when combined with stochastic algorithms, the network achieves comparable solutions in a shorter time frame.