Lye, Adolphus 
ORCID: 0000-0002-1803-8344, Prinja, Nawal and Patelli, Edoardo
(2022)
Towards a Robust Prediction of Material Properties by Artificial Intelligence and Probabilistic Methods.
In: International Conference on Topical Issues in Nuclear Installation Safety: Strengthening Safety of Evolutionary and Innovative Reactor Designs, 2022-10-18 - 2022-10-21, Vienna.
|
Text
IAEA-TIC 2022 Conference paper [Adolphus Lye].pdf - Author Accepted Manuscript Download (754kB) | Preview |
Abstract
The paper presents the results of a feasibility study aimed at combining probabilistic approaches for dealing with uncertainty with Artificial Intelligence (AI) technology for the prediction of the properties of materials used by the nuclear industry. This allows the AI tools to produce predictions with associated confidence, essential for the application of AI in safety critical systems. More specifically, this work involves predicting the Creep rupture and Tensile properties of a given type of steel material. To do so, a set of Artificial Neural Network (ANN) have been trained from relevant experimental data. However, the collected datasets are characterized by discontinuities and gaps in the data values. Furthermore, no information on its associated uncertainties is provided. To address these problems, a stochastic data-generating method is proposed which is used to enhance the dataset used to train the ANN models. From which, the Adaptive Bayesian Model Selection method is applied to obtain the corresponding probabilistic prediction with its associated confidence bounds. The results are well-validated against the given experimental data where the data is shown to fall within the prediction bounds. The approach has allowed for the improved accuracy of the prediction and making the model robust to bad data.
| Item Type: | Conference or Workshop Item (Unspecified) |
|---|---|
| Divisions: | Faculty of Science and Engineering > School of Engineering |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 02 Sep 2022 07:55 |
| Last Modified: | 18 Jan 2023 20:46 |
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3163062 |
Share
Share
