Performance and scalability of deep learning models trained on a hybrid supercomputer: Application in the prediction of the shear strength of slender RC beams

Abstract

Data-driven models employing artificial intelligence approaches have been increasingly utilized in structural analysis and design problems over the past two decades. The main applications involve the processing of datasets, which are gathered from experimentally derived records or obtained numerically, in order to develop closed-form formulae or numerical tools predicting quantities related to structural response and mechanical behaviour. Given that datasets are difficult to assemble due to the limited available information and the high cost entailed to enrich them, exhaustively processing the available data to produce the best possible prediction models is an essential task of particular interest. For specific applications, this exhaustive computing task involves large numbers of iterations performed to train detailed prediction models with large numbers of parameters. Despite the intense computational demands of such problems, limited research work exists on the scaling-up of the utilized algorithms on supercomputers. In this work, a distributed training and hyperparameter tuning algorithm is proposed for the modelling of the shear strength of slender beams without stirrups. The training dataset comprises results obtained from the detailed modelling and analysis of several beams with non-linear finite elements using the Reconan software. The results presented in this research work highlight the importance of optimally utilizing computational power for the solution of such problems. The developed computer code is available on GitHub.