Geometric Stability Classification: Datasets, Metamodels, and Adversarial Attacks

Many recent advances in machine learning have been motivated by classification problems. For example, classification methods are used to differentiate between spamand non-spamemails, identify hand written digits, and recognize the content of photos. For each application, a different model and model architecture will often perform best. Therefore, machine learning research has been enabled by readily available benchmark datasets. In particular, benchmark datasets have been used by researchers to demonstrate that novel methods can achieve high accuracy, and to demonstrate common vulnerabilities of classification methods to adversarial attacks. In the recent mechanics literature, there has been substantial interest in machine learning driven metamodels. Metamodels, or models of models, are appealing because once trained, they typically require orders of magnitude less compute time than full fidelity simulations. However, a better understanding of which machine learning methods and model architectures will perform best on mechanical data has been limited. Here we introduce an open source dataset BIC (Buckling Instability Classification) where a heterogeneous column is subject to a fixed level of applied displacement and is classified as either Stable or Unstable. In addition to introducing this benchmark dataset, we show baseline metamodel performance, and show two different types of adversarial attack. We anticipate that the open source BIC dataset will enable the future development of improved methods for classification problems in mechanics. (c) 2020 Elsevier Ltd. All rights reserved.

Automated summary

Recent advances in machine learning have been driven by classification problems and use of benchmark datasets. Metamodels in mechanics are of interest due to their efficiency in computation time. However, understanding which machine learning methods and model architectures perform best on mechanical data remains limited.