Department of Mechanical and Aerospace Engineering - PG Seminar - Unlocking Design Insights: Explainable Machine Learning in Engineering Optimization
Surrogate models, often considered a subset of supervised machine learning, have found widespread application in engineering design optimization. Besides optaining optimal solutions, it is equallly critical to gain knowledge and design insights that facilitate a deeper comprehension of the design problem. However, most surrogate models are typically treated as black boxes, lacking interpretability, which poses challenges for in-depth exploration of the relationship between input variables and objective functions. The introduction of explainable ML frameworks offers a solution by equipping us with a toolkit to elucidate the underlying mechanisms of these opaque models. This talk will discuss how to efficiently utilize the tools from explainable ML to uncover the inner mechanics of a high-dimensional model for extracting important insight. The central point of this talk revolves around the Shapley Additive Explanation (SHAP), which is now gaining traction in the field of engineering design. Due to its model-agnostic nature of SHAP, it can virtually be applied to a wide array of machine learning models and various types of input variables. The presentation will showcase several engineering applications including aerodynamic design, turbulence modeling, and the development of auxetic materials, from the perspective of both single and multiple objectives. In addition, this talk will discuss the exact computation of SHAP from a polynomial chaos expansion (PCE) surrogate model, allowing rapid computation of SHAP for engineering design analysis.
Pramudita Satria Palar is serving as an Assistant Professor at the Faculty of Mechanical and Aerospace Engineering, Bandung Institute of Technology (ITB), Indonesia. Previously, he held the position of Research Fellow at Tohoku University, Japan, and undertook a visiting research role at Leiden University. His research interests encompass computationally expensive optimization, surrogate models, statistical and machine learning, uncertainty quantification, and their application across engineering domains. He heavily employs surrogate models, artificial intelligence, and machine learning to advance engineering design and analysis.
Dr. Palar obtained his Ph.D. degree in Aeronautics and Astronautics from the University of Tokyo, Japan. During his doctoral research, he also visited Engineering Design Center at the University of Cambridge as a visiting Ph.D. student. In 2023, he had the honor of serving as an Invited Assistant Professor at ISAE-SUPAERO, France.