A machine learning methodology for the diagnosis of phase change material-based thermal management systems

Phase change materials (PCM) have received significant interest in various thermal energy storage and management applications due to their ample latent heat during the phase transition process. As PCM plays a vital role in these systems, knowledge of the state of the PCM is crucial for the sustained usage of the thermal management system. The energy absorbed by PCM as latent heat directly correlates with the average liquid fraction of the PCM; this can be used as a metric to monitor the thermal state of the system. Direct measurement of liquid fraction is quite challenging and is possible only through a thermal management system designed with a transparent material. This study proposes a machine learning-based diagnosis technique for a PCM-based thermal management system to predict the liquid fraction using surface temperature history. Recurrent neural networks (RNN) are chosen to predict the liquid fraction of PCM due to their non-linear and time-dependent nature. The data set required for the training of RNN is generated using numerical simulations. An RNN model is trained with a data set containing 345 samples which cover heat input types of constant, pulsating, random, Wiener, and discharging with corresponding temperatures as input and liquid fractions as target values of RNN. The results show that for all the heat inputs, the RNN can predict the temporal liquid fraction of PCM by showing a correlation up to 0.99 and RMSE less than 0.015 with the numerically obtained liquid fraction. Further, the RNN model takes a significantly lower computational time and power for predicting liquid fractions and can be deployed in real-life situations. Moreover, the study shows that challenging heat transfer problems are amenable to treatment with machine learning algorithms.