AI Trained to Predict Thresholds of 2D Ellipse Percolation Systems

Sammanfattning: Percolation theory is a relevant area of research in Nanotechnology because of its wide applications in nanoelectronics based on thin films of nanoparticles and composites, amongst others. In nanotechnology, systems are often explored through modelling and simulations. Thin films of the emerging low-dimensional nanomaterials, such as the 1D nanowires/nanotubes and 2D graphene, as well as their composites, can generally be simulated through a two dimensional percolation system of homogeneous or heterogeneous ellipses. The critical phenomena, i.e., the percolation threshold of the systems, is now obtained using the Monte Carlo simulation method, which, need extensive amounts of time. This project is an interdisciplinary one, wherein an attempt is made to use a certain amount of the data from the Monte Carlo simulations to train a machine learning model to predict the threshold of all the 2D ellipse systems with the maximum relative error < 10%, thus reducing the time taken when gathering the data. This project investigates different algorithms such as Linear Regression, Polynomial Regression, Multi-layer Perceptron Neural Networks, Random Forests, Extreme Gradient Boosted Trees, Support Vector Machines and K-Nearest Neighbours. Weaknesses in the results are identified and overcome by specific additional sample generation. Finally, a comparison is made between the algorithms marking the Multi-layer Perceptron and Extreme Gradient Boosted Trees as successful, with the Multi-layer Perceptron being the clear winner. The algorithm is successful within the defined 10% relative error, performing even better with all samples having relative prediction errors less than 7%. The model can be downloaded and used from https://github.com/NiravSurajlal/ PercolationAI.