[1A2] Using a machine learning forward model in an iterative depth profile reconstruction algorithm
J Hampton¹, A Fletcher¹, H Tesfalem¹, M Brown² and A Peyton¹
1University of Manchester, UK
2EDF Energy, UK
This paper compares the use of a finite element model and a machine learning algorithm for producing forward modelling data for use in eddy current non-destructive testing. This study considers a nuclear graphite inspection application; however, this method can also be applied in similar eddy current inversion problems and electromagnetic tomography reconstruction algorithms. We compare the performance of the algorithms against two prior estimates: that of a homogeneous bulk electrical conductivity and a prior from a machine learning direct inverse solver. For the comparison, we use 10 sets of synthetic data, with a noise profile derived from experimental data. We show that machine learning forward modelling techniques used within an optimisation algorithm are a feasible alternative to using a finite element model. There is a significant speed up of both the forward model and the inversion process when using a machine learning forward model, on average several orders of magnitude faster than using a finite element model. We then compare the machine learning and finite element forward models on a single sample of experimental data from an electrically conductive graphite block, of the type used in advanced gas-cooled reactors. Generally, the use of an iterative technique yielded a negligible reduction in error over a direct inverse solution, but the agreement with the solution from the finite element model is encouraging.
