Project Application Details

ARCHIE Project title:
Project PI: Civil and Environmental Engineering James Weir Building
Department: 17/05/2021
Address: Civil and Environmental Engineering, James Weir Building, open-source software, G1 1XJ
Relevant industrial collaborations: SEPA
Existing relevant KTP activities:
Does the PI require to run calculations? 8
Number of users involved in the project: 3 months
User 1 details:

DS username: dsb14145

Mr Euan Macdonald
euan.macdonald@strath.ac.uk
University of Strathclyde
Civil and Environmental Engineering
07950960103

Service level requested:
ARCHIE Project description:
Potential output of the project:

Network training, pruning and set assembly of an adaptive Baysian model selection process that is being implemented to make robust storm surge forecast predictions for the Firth of Clyde. The 3, 6 and 12 hour forecast prediction is made by training sets of neural networks on a number of pressure, wind stress and surge variables and then exploiting the variability of the training process across sets of artificial neural networks with randomized starting weights to quantify the uncertainty of the prediction. The model selection method uses Bayes theory (assuming uniform priors) to compute the posterior probability associated with each network in the set, in the form of a continuous distribution over the output domain. The robust prediction with confidence bounds are obtained by assigning weights to the individual network output according to the previously calculated posterior probabilities. Inefficient network architectures are pruned using optimal brain surgery method by assessing the second order derivative of the error function and removing the unimportant weights from the network architecture. The aim of simplifying the network complexity is to improve generalisation and increase the speed of further training.

Project start date: 17/05/2021
Duration of the project: 3 months
Total CPU time required (in hours): 20000
CPU hours required by typical job: 8
Approximate amount of data created per job (in GB):
Approximate amount of data created per project:
Funding Agency Grant Reference No: EP/R513349/1
pFact ID:
RKES ID:
Software required for the project: open-source software
Software required to be installed for the project: Tensorflow
Experience in using software package 1: 8
Software required to be installed for the project:
Experience in using software package 2: 8
Software required to be installed for the project:
Experience in using software package 3: 1
Training required:
Other training required:
Experience in using HPC: 8