Curing Cancer and Solving Climate Change: Practical Experimental Design with Normalizing Flows

Experimental design enables scientists to perform the most informative experiment under budget constraints, thus minimizing financial and human labor costs. Traditional approaches to experimental design are often complex and bound by restrictive assumptions. Here we present a simple modification to the forward Kullback-Liebler normalizing flow objective and show that the expected information gain (EIG) can be optimized jointly with the normalizing flow parameters. Addressing the challenges associated with optimizing binary designs, we adopt a probabilistic approach by learning a Bernoulli distribution for each design parameter. This results in a generalized algorithm that enables simple and flexible experimental design for a variety of realistic problems. Through two case studies – enhancing MRI medical imaging and improving seismic monitoring for CO2 plumes – we showcase the advantages of our method over random or manually chosen designs.