This project focused on reducing the sample cost of uncertainty quantification and surrogate modeling. I developed a backward-greedy design-of-experiments strategy for polynomial chaos expansion that starts from a larger candidate pool and removes the least informative samples, rather than following the standard forward-selection workflow.

Key contributions

• Designed a backward-greedy sampling procedure for design-of-experiments selection
• Improved both computational efficiency and approximation accuracy for polynomial-chaos-based surrogate construction
• Showed that the method reduces sensitivity to the particular optimality criterion used in practice

Impact

This work strengthened my foundation in uncertainty quantification and data-efficient surrogate design, which later informed my research in multi-fidelity scientific machine learning and other AI approaches for accelerating simulation.