Nested Sampling Strategy for Bayesian Design Space Characterization

Abstract Design space is a key concept in pharmaceutical quality by design, providing better understanding of manufacturing processes and enhancing regulatory flexibility. It is of paramount importance to develop computational techniques for providing quantitative representations of a design space, in accordance with the ICH Q8 guideline. The focus is on Bayesian approaches to design space characterization, which rely on a process model to determine a feasibility probability that is used for measuring reliability and risk. The paper presents three improvements over an existing nested sampling method: two-phase strategy with the first phase using a cheap sorting function based on nominal model parameters; dynamic sampling strategy to refine the target design space; and vectorization to evaluate costly functions in parallel. These improvements are implemented as part of the python package DEUS and demonstrated on an industrial case study.