Neural Networks based Conformal Prediction for Pipeline Structural Response

The widespread use of machine learning models has achieved considerable success across various domains. Nevertheless, their deployment in safety-critical systems can result in catastrophic consequences if uncertainties are not handled properly. This study is concerned with the simulation of the physical response of a subsea pipeline when it is hooked by an anchor. Predicting this response is crucial for risk assessment, however, it is computationally unfeasible to run a significant amount of input sets to compute the probability of failure of the system. Therefore, the use of a surrogate model becomes essential. In this context, a surrogate model is a machine learning model trained on data from a physicsbased simulation. This is achieved by neural network based surrogate models, as they are capable of modelling complex relationships and provide greater accuracy than other machine learning models in many use cases. However, to ensure the safe use of these models, it is important to understand the uncertainty associated with their predictions. Therefore, we apply the conformal prediction framework to provide valid prediction intervals and improve the uncertainty quantification of the neural network models. In order to create adaptive conformal prediction intervals, we employ multilayer perceptron neural network models that provide uncertainty estimates through both the Monte Carlo dropout technique and treating the output as a Gaussian distribution, with the neural network providing estimates for both mean and variance. The conformal prediction procedure improves the uncertainty estimation of uncalibrated models and guarantees new test samples are within the predicted intervals with the corresponding selected confidence level.