Research project

Knowledge Transfer Secondment: Bayesian learning with complex computer models

Project overview

Pharmaceutical and chemical sciences regularly face the challenge of learning about complex processes using data collected from statistically designed experiments. When these data are too complex to be modelled by low-order polynomials and insufficient mechanistic knowledge exists to build physical models, more sophisticated statistical learning methods are required.

Bayesian learning, using data-driven nonparametric models, starts from prior beliefs, formulated in a prior distribution, and uses collected data to update uncertainty about the system under study via Bayes theorem. Use of these probabilistic models, which provide realistic and accurate uncertainty quantification, can lead to better exploitation of industrial and business data, and hence more reliable science, lower development costs and faster time to market.

The most common tool for nonparametric Bayesian learning is the Gaussian process. Much research effort and application resource has been expended in developing and demonstrating efficient modelling methodologies for a variety of data. However, there is considerably less research into, and fewer applications of, the design of experiments for Bayesian learning. As data collection in many industrial processes, particularly in pharmaceutical science, is expensive and time consuming, collection of the minimum amount of most informative data is vital to successful practical exploitation of Bayesian learning.

This project will look at industrial applications of Gaussian process-based Bayesian learning supplier by GlaxoSmithKline, including from chemical development and manufacturing and computational chemistry.

Staff

Lead researchers

Professor Dave Woods

Professor of Statistics

Research interests

  • Design of experiments
  • Bayesian statistics
  • Statistical computing
Connect with Dave

Collaborating research institutes, centres and groups

Research outputs