Simulation and artificial intelligent methodologies for end-to-end bio-pharmaceutical manufacturing and supply chain risk management

Driven by the critical challenges in the bio-pharmaceutical industry, including complexity, high variability, and lengthy lead time, we create new simulation and artificial intelligent methodologies to improve the end-to-end bioprocess understanding, guide the decision making, facilitate the develop...

Full description

Bibliographic Details
Published:
Online Access:http://hdl.handle.net/2047/D20398282
Description
Summary:Driven by the critical challenges in the bio-pharmaceutical industry, including complexity, high variability, and lengthy lead time, we create new simulation and artificial intelligent methodologies to improve the end-to-end bioprocess understanding, guide the decision making, facilitate the development of modular, flexible, intensified, reliable and automated biomanufacturing supply chain. First, we construct a stochastic simulation for biomanufacturing process to control the key sources of uncertainty leading to batch-to-batch production variation. It incorporates the physical chemical interactions and dynamics. This simulation model can be used to guide biomanufacturing risk management and coherent operational decision making. Second, to ensure the drug safety, delivery reliability, transparency, responsiveness, and data integrity, we introduce a blockchain enabled interoperability framework with reputation-based Proof-of-Authority smart contract and state sharding, which can efficiently utilize the supply chain surveillance resources and simultaneously process the jobs from different areas. Then, a simulation-based platform is developed for the biopharmaceutical supply chain risk management. Third, to ensure the digital twin faithfully represents the real biopharmaceutical systems; and balance computational time and calibration accuracy, we develop a new Bayesian sequential design of experiments for simulation calibration. It can simultaneously support the real-time digital twin calibration and guide the optimal decision making. Fourth, we propose a quantile estimator by pooling detailed simulation trajectories from parallel computing, that can efficiently use the computational resources and accurately assess the risk behaviors. Then, we introduce a distributional metamodel which can model a sequence of percentile surfaces and improve the system risk performance prediction accuracy and efficiency. Further, we introduce a new probabilistic knowledge graph characterizing the end-to-end bioprocess spatial-temporal causal interdependencies. Systematic risk and sensitivity analyses are proposed to quickly identify and remove the bottlenecks, guide the process specifications and risk control, and accelerate quality-by-design. Fifth, to support the biomanufacturing automation, we introduce a bioprocess model-based reinforcement learning accounting for model risk, which can simultaneously support bioprocess online learning and guide the customized stopping policy for multi-phase fermentation process. By conducting structural and sensitivity analyses and comprehensive empirical study, we investigate the impact of model risk on the optimal policy and provide insightful guidance on fermentation process control.