Predicting the Thermodynamic Properties of Fluids Using Neural Networks

Machines learning techniques have proved useful in various fields including business, medicine, transportation and fields in the physical sciences such as particle physics. This research aims to gauge the effectiveness of deep learning techniques in modeling thermodynamic systems. Specifically, the work optimizes a multi-layer neural network to quantitatively fit the van der Waals equation of state for single component system and mixtures. Multi-layer neural networks connections each have weights and biases which represent their importance within the network. These weights and biases are adjusted with the backpropagation algorithm to create accurate predictions. The research applies these principles to predict the pressure given by van der Waals equation given volume, temperature, and number of molecules as inputs. The van der Waals model is a modification of the Ideal Gas law. The model was conceived in 1873 by Johannes van der Waals to more accurately describe the qualitative behavior of fluids within a mathematical model. The model is described by van der Waals equations: which can also be extended to mixtures. By training a neural network to fit to van der Waals model, we plan to understand how the choice of hyperparameters (number of layers, number of neurons etc.) affects the accuracy of the predictions. Future work aims to apply similar techniques to predict the equations of state for various fluids whose thermodynamic properties have no accurate mathematical description.