Bayesian Optimal Experimental Design
Designing experiments for inference and decisions under uncertainty
My research asks a simple question:
Which experiment should we run next, and why?
I study Bayesian optimal experimental design (BOED) for scientific systems where data are expensive, models are uncertain, and the ultimate goal is often a decision rather than parameter estimation alone.
Classical BOED is powerful but difficult to apply in modern scientific settings. My work targets four practical bottlenecks:
| Bottleneck | My direction |
|---|---|
| Expensive simulations | Neural operator surrogates for scalable BOED |
| Misspecified priors | Robust design under prior uncertainty |
| High-dimensional parameters | Diffusion- and flow-based posterior sampling |
| Downstream decisions | Decision-focused experimental design |
Together, these projects aim to make BOED practical for large-scale inverse problems, PDE-governed systems, and scientific decision-making.
Project directions
Scalable BOED with Neural Operators
Large-scale BOED requires repeated forward-model evaluations, which quickly become prohibitive for PDE-governed systems. I use neural operator surrogates to make Bayesian design tractable at scale.
Robust and Decision-Focused BOED
Standard information gain can fail when the prior is misspecified or when not all parameter uncertainty is relevant to the final decision. I develop robust and goal-aware design objectives that target only the information that matters.
Generative Priors for High-Dimensional BOED
For high-dimensional inverse problems, posteriors are often non-Gaussian and difficult to sample. I use diffusion and flow-based generative models to enable scalable posterior sampling and information-gain estimation.
LLM-Elicited Priors
Specifying a prior is often the hardest step in Bayesian inference. I study how large language models can help elicit structured prior knowledge for experimental design and decision-making.