Uncertainty Quantification
Pharmaceutical and bioprocess models are inherently uncertain: processes are often stochastic, parameters are estimated from noisy data, and the models themselves are approximations built on many assumptions. Ignoring this uncertainty leads to overconfident predictions and unreliable decisions. My research addresses uncertainty end-to-end — from reducing it experimentally, to quantifying and propagating it through models, to making decisions that explicitly account for it.
Mitigate
Uncertainty can be reduced before it is a problem by designing experiments that are maximally informative. I develop model-based optimal experimental design (OED) methods that select sampling times, input trajectories, and measurement types to minimise parametric uncertainty in the resulting model. This is especially important in pharmaceutical and bioprocess applications where experiments are expensive.
Quantify
Once a model is built, I quantify how uncertain its parameters are using Bayesian parameter estimation, global sensitivity analysis (GSA), and practical identifiability analysis. GSA tells us which parameters matter most; identifiability analysis tells us which can be estimated reliably from available data.
Propagate
Parametric uncertainty must be propagated through nonlinear models to obtain uncertain predictions. I use polynomial chaos expansions (PCE) and Monte Carlo methods, benchmarking their efficiency and accuracy across different process models — from population balance models for milling to kinetic models for fermentation.
Decide
Ultimately, uncertainty must inform decisions. I develop robust optimisation, chance-constrained optimisation, and risk-averse model predictive control (MPC) strategies that account for model uncertainty when computing optimal operating conditions or control actions.