Abstract

Statistical modeling techniques are gaining popularity as surrogate models to mechanistic modeling such as finite element analyses (FEA). The surrogate aims to map inputs to outputs while the FEA estimates the solution using the governing equations. Previous work has shown that machine learning algorithms have modeled highly non-linear data. Additionally, transfer learning has been a method applied in computer vision and natural language processing to reuse common knowledge learnt from a particular task to a different, but related task. However, the application of such implementations to engineering simulations is only starting to emerge. The objective of this paper is to demonstrate the relevance of transfer learning using simulation data in order to make real-time accurate life predictions, prepare overhaul schedules and detect anomalies. We developed a fully connected deep neural network for this regression task of fitting data from a finite element model. Subsequently, we applied transfer learning by utilizing scarce field data from the real model, which is alike in geometry, but varying in physical and material properties. In-house tools were deployed to efficiently generate 27,000 source acceleration-stress profiles to train the network as a stress predictor based on acceleration. Furthermore, 6000 target acceleration-stress profiles were computed to adjust the predictor to the target model. Popular python-based machine learning tools were employed to explore a hyperparameter space. This analysis facilitated in recognizing key learning patterns, apprehending knowledge transfer techniques involving freezing, retraining and adding layers to the base neural network and finally in identifying the optimal neural network configuration for the target task. Excellent convergence in loss functions were calculated giving us confidence in the training and transfer learning processes. This research showed that the tuned model is capable of transferring knowledge gained from fitting a large dataset to a different and smaller dataset, thus creating an accurate and useful digital twin from FEA.