Constitutive relations are key to continuum modeling of mechanical response of granular media relevant to geotechnical engineering. Conventional approaches in the development of constitutive models have largely been knowledge and experience based. Due to the highly nonlinear and path-dependent characteristics of granular materials, conventional constitutive relations for granular media are frequently loading-condition dependent and problem-specific and often contain considerable phenomenological assumptions. A new, data-driven method is proposed to derive the complex constitutive relationship of granular media based on latest advances in data science on machine learning and deep learning.

In this research, long short-term memory (LSTM) network, a variant of recurrent neural net...[ Read more ]

Constitutive relations are key to continuum modeling of mechanical response of granular media relevant to geotechnical engineering. Conventional approaches in the development of constitutive models have largely been knowledge and experience based. Due to the highly nonlinear and path-dependent characteristics of granular materials, conventional constitutive relations for granular media are frequently loading-condition dependent and problem-specific and often contain considerable phenomenological assumptions. A new, data-driven method is proposed to derive the complex constitutive relationship of granular media based on latest advances in data science on machine learning and deep learning.

In this research, long short-term memory (LSTM) network, a variant of recurrent neural network for sequential problems, is adopted for bridging the gap between stress and strain of granular media. The history information including stress and strain are fed into the model as input features step by step. The model is then trained to learn the relationship between different steps and finally obtains the capability to predict the stress state as well as other qualities of the media at next time step. This LSTM model is first built, trained and evaluated based on simple shear test data generated from high-fidelity DEM simulations. The predictions from both this LSTM network model and a conventional feedforward neural network model are compared against the given dataset. The LSTM model is found to perform marginally better than the feedforward neural network model in most predictions. The performance of this LSTM model is further assessed on the experimental dataset from laboratory tests of both monotonic and cyclic loading on fine sand. It shows a remarkable predictive ability in capturing the complex constitutive responses of granular media under various loading conditions. A multiscale modeling framework is finally proposed to couple the material point method (MPM) and the machine learning model together. The machine learning model is firstly trained on offline generated dataset and is then embedded into the coupled framework. The newly developed data-driven multiscale modeling method is benchmarked and demonstrated for its robust predictive powder for granular media.

Sequential machine learning models have the potential to capture the complex constitutive characteristics of granular media under various conditions and it is possible to build constitutive models purely with sufficient data rather than empirical assumptions.