PhD Oral Exam - Khaled Salem Al Hammad, Mechanical Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Water droplet erosion (WDE) significantly compromises the durability and performance of critical components in aerospace, power generation, and wind energy industries. Accurate prediction of WDE behavior, particularly the incubation period and maximum erosion rate, is vital for effective material selection and component design. Traditional empirical and analytical models often struggle to capture the complex, non-linear interactions of material properties and impact conditions, limiting their predictive accuracy and generalizability. To overcome the limitations of existing models, this thesis introduces a comprehensive framework that leverages machine learning (ML) techniques for robust and interpretable predictions of WDE. Through a series of interconnected investigations, this work develops and validates an ML approach to accurately predict the WDE incubation period and maximum erosion rate (Re), demonstrating the effectiveness of this methodology. This includes evaluating various ML algorithms and investigating the impact of data transformation techniques on prediction accuracy. Findings revealed that statistical data transformations, such as Box-Cox and Yeo-Johnson, substantially enhanced model performance, with a linear regression model achieving an R² value above 90%. Feature importance analysis, particularly through SHAP (SHapley Additive exPlanations) values, identified impact velocity and surface hardness as the most influential factors, providing valuable physical insights into the erosion mechanisms. To predict the maximum erosion rate, this research addresses the challenge of limited experimental datasets in WDE studies. It showcases the transformative impact of data augmentation strategies, such as Synthetic Minority Over-sampling Technique for Regression (SMOTER). These strategies enabled highly accurate predictions, with a K-Nearest Neighbors (KNN) model achieving an R² exceeding 95% on training and testing data and maintaining that accuracy on unseen validation data, despite the small initial datasets. This independently validates the effectiveness of data augmentation in enhancing predictive accuracy and generalization for maximum erosion rate prediction. This study also explores the interpretability and generalization of data-driven models for predicting the WDE incubation period. Traditional methods, such as Multiple Linear Regression (MLR) with dimensionless analysis and Principal Component Analysis combined with a Neural Network (PCA-NN), are compared with advanced tabular architectures such as the Feature Tokenizer Transformer (FT-Transformer), Tabular Neural Network (TabNet), Self-Attention and Intersample (SAINT), and Tabular Prior-Data Fitted Network (TabPFN). While advanced transformer-based models achieve strong predictive performance on training and testing datasets, their generalization varies when applied to an unseen material dataset. Notably, TabPFN and SAINT demonstrated more robust extrapolative capabilities, underscoring the importance of balancing accuracy, interpretability, and generalization.

A significant contribution of this work is developing a physics-informed model for the WDE incubation period (No) using Buckingham-Π dimensional analysis. This framework incorporates key parameters, including droplet impact velocity, droplet diameter, material yield strength, fracture toughness, speed of sound in both material and water, material density, water density, Vickers hardness, and Poisson’s ratio. The resulting dimensionless power-law model highlights the effectiveness of physics-informed approaches in predicting WDE incubation period and capturing the complex interplay among impact conditions and material properties.

Collectively, this thesis bridges the gap between traditional empirical approaches and modern ML-based techniques, providing robust, accurate, and interpretable predictive models for WDE. The findings contribute to improved material selection, component design, and maintenance strategies in systems prone to WDE by offering a data-driven framework that enhances both predictive capability and mechanistic understanding.