A machine learning-based multi-scale computational framework for granular materials

Deep learning‐accelerated multiscale approach for granular material modeling

“…Notably, in some existing studies, the tangent operator was also involved in the training process. For example, an additional model for predicting the tangent operator was constructed in a relevant study, 28 whereas both stress and tangent operator were included in the output of the model in other studies 29,30 . However, the inclusion of a tangent operator in the training process significantly raises the computational cost.…”

“…The maximum error was observed inside the shear band, whereas the errors in the other regions were rather small, which is consistent with findings reported in the literature. 29,31,49 This is because the strain level in the shear band is significantly higher than that outside, and the testing data can fall beyond the range of the training data. 50 Although the DLM optimally performs interpolation, it often exhibits a weaker extrapolation ability.…”

Deep active learning for constitutive modelling of granular materials: From representative volume elements to implicit finite element modelling

“…In contrast, MLP is more efficient but has to introduce extra internal variables to capture the loading-path dependence of strain-stress responses. One option is by using the accumulation of absolute strain increments φ to encode the loading history (Guan et al, 2022;Huang et al, 2020):…”

“…Unlike the phenomenological constitutive theory and multiscale modelling, data-driven models do not require parameter calibration and phenomenological assumptions, neither do they request unaffordable computational resources to infer stress responses from strain paths. Although it is not new to apply neural networks to model the stress-strain relations of concrete and sands (Ellis et al, 1995;Ghaboussi et al, 1991;Ghaboussi and Sidarta, 1998), the revolutionary development of deep learning over recent years re-inspires extensive explorations in data-driven constitutive models (Guan et al, 2023;Ibragimova et al, 2022;Ibragimova et al, 2021;Jordan et al, 2020;Tancogne-Dejean et al, 2021). For example, and developed reinforcement learning and game theory-based deep learning models for the constitutive modelling of granular materials.…”

Data-driven strain–stress modelling of granular materials via temporal convolution neural network

“…Particularly, recurrent neural networks (RNNs) (Qu et al, 2021a,b;Ma et al, 2022) are generally applied in describing constitutive laws for granular materials and can acquire good prediction results (Gorji et al, 2020;Mozaffar et al, 2019). The trained model has also been adopted to simulate macroscopic problems within the coupled finite element and machine learning modelling framework (Shaoheng Guan et al, 2022).…”