Bayesian Optimization of Deep Learning Algorithms in Sorption Processes

The use of machine learning algorithms often involves careful tuning of the learning parameters and hyperparameters of the model. Unfortunately, this tuning is often very difficult and requires a lot of experience. Therefore, there is a high demand for automated approaches that can optimize the performance of any learning algorithm for a given problem.

Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU) are among the most popular deep learning models currently in use. The performance of LSTM, BiLSTM and GRU is highly dependent on the selection of several hyperparameters that must be carefully selected to obtain good results. In this work, we consider this problem through the framework of Bayesian optimization.

The article presents deep learning methods used to predict the mass of an adsorption bed in a fluidized bed. The purpose of using this type of bed is to improve the efficiency of adsorption cooling systems by increasing heat and mass transfer through the use of fluidization. The article presents numerical studies on mass prediction using the above-mentioned algorithms for silica gel as a sorbent. In selected neural networks, the following hyperparameters were optimized: number of hidden layers and number of neurons in each layer.

The described research is part of the work planned in the project No. 2018/29/B/ST8/00442, "Research on methods of intensifying the sorption process in modified adsorbent deposits", financed by the National Science Center, Poland.