Robust artificial neural network for reliability and sensitivity analyses of complex non-linear systems.

Area of Science:

• Computational Science and Engineering
• Machine Learning and Artificial Intelligence
• Nuclear Engineering and Safety

Background:

• Artificial Neural Networks (ANNs) reduce computational cost for uncertainty quantification, reliability, and sensitivity analyses.
• Random initialization in ANNs leads to performance variations and uncertainty in model selection.
• R-squared (R²) for cross-validation can bias predictions and inadequately assess model performance.

Purpose of the Study:

• To propose an approach for enhancing the robustness of Artificial Neural Network (ANN) predictions.
• To improve the reliability and sensitivity analyses by addressing prediction biases and uncertainties.
• To quantify prediction uncertainties using confidence intervals.

Main Methods:

• A systematic combination of identically trained ANNs is employed.
• The approach integrates the Bayesian framework with model averaging techniques.
• Two synthetic numerical examples and a UK nuclear effluent treatment plant simulation model were used for demonstration.

Main Results:

• The proposed method enhances prediction robustness by systematically combining multiple ANNs.
• Uncertainties in predictions are effectively quantified using confidence intervals.
• The approach was successfully applied to reliability and sensitivity analyses of a nuclear process simulation model.

Conclusions:

• The proposed Bayesian framework and model averaging approach significantly improves ANN prediction robustness.
• This method provides a reliable way to perform uncertainty quantification, reliability, and sensitivity analyses.
• The approach is applicable to complex, black-box process simulation models in nuclear engineering.