A multiscale multi-permeability poroplasticity model linked by recursive homogenizations and deep learning
By Kun Wang, WaiChing Sun. 2018
Abstract Many geological materials, such as shale, mudstone, carbonate rock, limestone and rock salt are multi-porosity porous media in which pores of different scales may co-exist in the host matrix. When fractures propagate in these multi-porosity materials, these pores may enlarge and coalesce and therefore change the magnitude and the principal directions of the effective permeability tensors. The pore-fluid inside the cracks and the pores of host matrix may interact and exchange fluid mass, but the difference in hydraulic properties of these pores often means that a single homogenized effective permeability tensor field is insufficient to characterize the evolving hydraulic properties of these materials at smaller time scale. Furthermore, the complexity of the hydro-mechanical coupling process and the induced mechanical and hydraulic anisotropy originated from the micro-fracture and plasticity at grain scale also makes it difficult to propose, implement and validate separated macroscopic constitutive laws for numerical simulations. This article presents a hybrid data-driven method designed to capture the multiscale hydro-mechanical coupling effect of porous media with pores of various different sizes. At each scale, data-driven models generated from supervised machine learning are hybridized with classical constitutive laws in a directed graph that represents the numerical models. By using sub-scale simulations to generate database to train material models, an offline homogenization procedure is used to replace the up-scaling procedure to generate cohesive laws for localized physical discontinuities at both grain and specimen scales. Through a proper homogenization procedure that preserves spatial length scales, the proposed method enables field-scale simulations to gather insights from meso-scale and grain-scale micro-structural attributes. This method is proven to be much more computationally efficient than the classical DEM–FEM or FEM2 approach while at the same time more robust and flexible than the classical surrogate modeling approach. Due to the usage of bridging-scale technique, the proposed model may provide multiple opportunities to incorporate different types of simulations and experimental data across different length scales for machine learning. Numerical issues will also be discussed.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
Citations in the corpus
- Physics informed deep learning for computational elastodynamics without labeled data (2020, background): “…into two streams: (1) the use of experimental or computationally-generated data to create coarse-graining reduced-fidelity or surrogate models [3–7], and (2) physics-informed neural network (PINN) for modeling the solution of partial differential equations (PDEs) that govern the behavior of…”
- Physics-Informed Multi-LSTM Networks for Metamodeling of Nonlinear Structures (2020, background): “, convolutional neural network (CNN) [37] and recurrent neural network (RNN) [38, 39]) is a promising approach to establish metamodels for fast prediction of time history response of dynamical systems [36, 40–42] and material constitutive modeling [43, 44].”
- Meta-modeling game for deriving theoretical-consistent, micro-structural-based traction-separation laws via deep reinforcement learning (2018, methodology): “…(e.g. identifying material parameters (e.g. Wang et al. [2016], Liu et al. [2016a]) or adjusting weights of neurons in recurrent neural networks (e.g. Lefik and Schrefler [2002], Wang and Sun [2018]) is mutually exclusive to the testing dataset used to evaluate the quality of blind predictions.”
- A Physics-Constrained Data-Driven Approach Based on Locally Convex Reconstruction for Noisy Database (2019, background): “In contrast, most other data-driven methods for solid or fluid simulation directly construct machine learning surrogates to link the input-output relation with approximated physics laws [17, 24, 33] or with the constitutive models replaced by supervised learning models such as neural networks [12-14”
- Geometric deep learning for computational mechanics Part I: Anisotropic Hyperelasticity (2020, methodology): “One major advantage of the hybrid (Frankel et al., 2019) or grah-based training (Wang and Sun, 2018) is the ability to generalize the forward prediction ranges by introducing RVE data as an additional initial input.”