Online Bayesian Re-design of Parallel Experiments based on Asynchronous Posterior Sampling

MARTIN F. LUNA; M. NICOLÁS CRUZ BOURNAZOU; ERNESTO MARTÍNEZ

Toulouse

Simposio; 32nd European Symposium on Computer Aided Process Engineering (ESCAPE32); 2022

European Federation of Chemical Engineering

High-throughput robotic platforms boost gathering informative data sets to support bioprocess development by resorting to on-line redesign of asynchronous parallel experiments. Due to significant uncertainty in both structure and parameters of mathematical models of bioreactors during early developmental stages, a probabilistic Bayesian approach is proposed. A novel algorithm that combines Asynchronous Posterior Sampling (APS) of model parameter distributions with chance-constrained optimization is used to bias data sampling depending on the modeling goal. As new data are available, model parameter distributions are updated using variational Bayesian inference. Myopic posterior sampling is then used online for purposefully changing cultivation conditions in parallel experiments. The proposed approach is based on a probabilistic macroscopic model, whereas the modelling goal is specified by integrating domain expertise and preferences via a reward function. A case study related to Escherichia coli expressing a desired product is used to demonstrate that a trade-off between improving parametric precision and biasing data gathering towards bioprocess optimization is achieved. Results obtained are encouraging for autonomous operation of robotic platforms.