INTERPRETABLE MACHINE LEARNING MODELS FOR AEROSPACE ENGINEERING APPLICATIONS

• Main Author: Stevenson, Rafael
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70018
• Institution: Institut Teknologi Bandung

As data-driven science and engineering by machine learning method is currently emerging as a popular method due to its outstanding performance in physics modeling tasks, machine learning models are widely used to perform predictive modeling, especially in aerospace engineering and fluid science field where the phenomena mostly involve high-dimensional and non-linearity. However, popular models such as artificial neural network usually need a large amount of training data to perform well whereas unfortunately labeled data in the aerospace engineering field which are usually obtained from computational simulation or experiments could be very expensive to obtain. Therefore, this thesis demonstrated how a particular machine learning model called gene expression programming could have the capabilities to perform better than other popularly used machine learning models in less training data environments. Also, another emerging challenge posed by using machine learning models is their degree of interpretability. In military and aerospace engineering field, applying machine learning model for real-life cases require justifiable reasons behind the prediction where data-driven models have to be certifiable and verifiable for safety reasons. Addressing those challenges in data-driven machine learning modeling, this research demonstrates how model-specific interpretation could be performed to explain white-box interpretable machine learning models and also even explain less interpretable black-box models by model-agnostic interpretation methods particularly for case studies related to the aerospace engineering field.