Machine Learning approaches for robust analysis of complex systems

Overview

Decommissioning of legacy nuclear facilities from decades old military and civil programmes represents a significant challenge to the UK. Due to the fact that uncertainties are not underpinning, the potential to optimise decommissioning strategies could save the UK tax payer many £100’s millions of pounds over the life time of decommissioning of a single large facility, such as the MAGNOX reprocessing facilities at Sellafield.

Determining the optimal strategy for nuclear decommissioning is dependent on generating realistic data to represent future scenarios. This is a particular problem in effluent management on nuclear sites where historic data representing the effluent challenge generated during normal operations under largely steady state conditions provides a poor representation of challenges that will occur during decommissioning. Effluent feeds consist of radioactive ions that must be abated and competing ions that affect the abatement performance. The problem is exacerbated by the fact that abatement is understood to respond in a non-linear fashion to stochastic variation of the feed challenge. The challenge is to identify methods to generate artificial data that could reflect changes brought upon by decommissioning but retain the inherent variability that is currently observed. This would require the development of methods to establish the credibility of this data by determining its validity and veracity. Establishing credibility of the approach is a key step in ensuring that the methods can be adopted within decommissioning programmes to realise the benefits as outlined.

The applicant should have a very strong mathematical background, computational skills, and knowledge in probability theory and mathematical statistics.

Project Description

The purpose of this research is to provide a novel and general approach based on machine learning technique and aritficial intelligence  for the evaluation of the risk associated with radionuclide discharge on nuclear sites during decommissioning and able to overcome the limitations of the current approaches. Effluent feeds consist of radioactive ions that must be abated and competing ions that affect the abatement performance. The problem is exacerbated by the fact that abatement is understood to respond in a non-linear fashion to stochastic variation of the feed challenge. This understanding is obtained from mechanistically based models of abatement as a function of chemical reactions and physical processes including settling, filtration and flow. The challenge is to identify methods to generate artificial data that could reflect changes brought upon by decommissioning but retain the inherent variability that is currently observed.

This would require the development of methods to establish the credibility of this data by determining its validity and veracity. A key part of the work will be to demonstrate to stakeholders that artificial plant data generated in this way is credible. .

Many sources of epistemic and aleatory uncertainties are present in the estimation of the ground acceleration capacity. The project will explore the possibility to use advance methods for uncertainty characterization and quantification including extreme value statistics, Bayesian and generalized probabilistic methods. The project will involve working closely with National Nuclear Laboratory and Sellafield Ltd working with plant data and in collaboration with process engineers and other industry stakeholders. The project will include a placement of a minimum of 3 months to NNL facilities at Risley.

Aims and Outcomes

  • Proper treatment of uncertainty

  • Increase the reliability of developed method

  • Development of an realistic simulator able to predict the future conditions on nuclear sites during decommissioning

  • Provide a simple yet robust simulation tool able to deal with uncertainty

     

    For informal information please send your CV to edoardo.patelli@liverpool.ac.uk

Candidate Requirements

This post is suitable for applicants who have graduated with a first class degree (or equivalent) in Engineering, Mathematics, Physics or a related/relevant technical discipline. Candidates should have a good computational skills and a strong interest in modelling and simulation. Experience of using Matlab would be advantageous but not a necessity.   

For further details please send a copy of your curriculum vitae to Dr. Edoardo Patelli (edoardo.patelli@liverpool.ac.uk)

Centre for Doctoral Training in Quantification and Management of Risk & Uncertainty in Complex Systems & Environments is funded by EPSRC and ESRC.