Solving Ordinary Differential Equations and Systems using Neural Network Methods

Abstract: The applications of differential equations are many. However, many differential equations modelling real-world scenarios are very complex and it can be of great difficulty to find an exact solution if one even exists. Thus, it is of importance to be able to approximate solutions of differential equations. Here, a method using neural networks is explored and its performance is compared to that of a numerical method. To illustrate the method, two first order, two second order and two first order systems of ordinary differential equations are explored. The systems are the Lotka-Volterra system and the SEIR (Susceptible, Exposed, Infected, Removed) epidemiological model. The first four examples have exact solutions to compare to and the observations are then used as a basis when discussing the results of the systems. The results of the thesis show that while the neural network method takes longer to deliver an approximation, it continuously gives better approximations than the implicit Euler method used for comparison. The main contribution of this thesis is the comparison done of the performances of the neural network method and the implicit Euler method.