A design centering methodology for probabilistic design space

Abstract

The use of mathematical models for design space characterization has become commonplace in pharmaceutical quality-by-design, providing a systematic risk-based approach to assurance of quality. This paper presents a methodology to complement sampling algorithms by computing the largest box inscribed within a given probabilistic design space at a desired reliability level. Such an encoding of the samples yields an operational envelope that can be conveniently communicated to process operators as independent ranges in process parameters. The first step involves training a feed-forward multi-layer perceptron as a surrogate of the sampled probability map. This surrogate is then embedded into a design centering problem, formulated as a semi-infinite program and solved using a cutting-plane algorithm. Effectiveness and computational tractability are demonstrated on the case study of a batch reactor with two critical process parameters.