Physics-informed Interpretable Machine Learning

Midshipman Researcher(s): 2/C Wesley Nourachi

Adviser(s): Professor Kevin McIlhany

Poster #87

Traditional solutions to the diffusion equation are computationally expensive when boundary conditions change with time. We have formulated an approximate matrix solution which is computationally cheaper than traditional approaches and can be implemented in a neural network to find a higher-order solution. Additionally, LiDAR point cloud data of urban infrastructure is one such case where incredibly large datasets need to be reduced to much lower dimensionality while still maintaining all the information contained therein. To that end, we have determined that neural networks are capable of performing principle component analysis, although generally with less accuracy for isolated structures.

Full Size Physics #87