A comparative study of machine learning approaches for modeling concrete failure surfaces. (February 2018)

• Record Type: Journal Article
• Authors: Reuter, Uwe; Sultan, Ahmad; Reischl, Dirk S.
• Is Part Of: Advances in engineering software. Volume 116(2018)
• Page Start: 67
• Page End: 79
• Publication Date: 2018-02
• Subjects: Artificial neural network -- Support vector machine -- Support vector regression -- Concrete failure surface; Computer-aided engineering -- Periodicals; Engineering -- Computer programs -- Periodicals; Engineering -- Software -- Periodicals; Periodicals; 620.0028553
• DOI: 10.1016/j.advengsoft.2017.11.006
• Languages: English
• ISSNs: 0965-9978

Abstract

Highlights: A general shape function, which is able to generate types of surfaces that cover most of the hypothesis of concrete failure criterion, is introduced. Realizations of this general shape function with different noise levels emulating experimental errors are used for verification purposes of the machine learning approaches. Artificial neural networks, support vector machines, and support vector regression are adapted to model the failure surface of C25 concrete starting from 88 experimental tests. These model-free approaches are independent from of any predefined models and eliminate the need of new models for new concrete types. Abstract: This study introduces an enhanced approach for concrete failure criterion, which is strongly needed for a realistic simulation of concrete behavior, by employing machine learning approaches instead of the traditional models of failure surfaces. Since the shape of concrete failure surfaces is not exactly known, a general shape function for verification purposes of the machine learning approaches is introduced. Artificial neural networks, support vector machines, and support vector regression are adapted to model realizations of this general shape function with different noise levels. After the successful fitting of these surfaces, the algorithms are employed to model the failure surface of C25 concrete starting from 88 experimental tests. The three approaches are able to fit the experimental data with low error and are compared to one another. Drucker–Prager and Bresler–Pister surfaces are solved for the same experimental data and compared with the support vector regression surface. The main advantage of machine learning approaches is that they are model-free approaches which eliminate the need of new models for new concrete types.