Transfer learning of deep material network for seamless structure–property predictions

Liu, Zeliang; Wu, C. T.; Koishi, Masataka

doi:10.1007/s00466-019-01704-4

Cited by 48 publications

(31 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We consider DMN [1,2] as a blend of both approaches: it finds a reduced-order representation of the DNS RVE model while the training only requires stress-strain data. Moreover, the transfer learning approach of DMN for creating a unified microstructure database is proposed in [53] and applied to short-fiber reinforced composites in [54]. The thermodynamic consistency of DMN has also been studied in [55].…”

Section: Motivation and Backgroundmentioning

confidence: 99%

“…Although not discussed in this paper, the network interpolation from transfer learning is also an appealing feature of DMN. In [53,54], databases of multiple microstructures are unified to cover the design space of composites. The learned database can then be applied to concurrent multiscale simulations with local microstructure variations resulting from the manufacturing processes, such as the injection molding of short fiber-reinforced composites and the metallic additive manufacturing.…”

Section: Appendix a Machine Learning Of Dmn: Data Generation And Opti...mentioning

confidence: 99%

See 1 more Smart Citation

Cell division in deep material networks applied to multiscale strain localization modeling

Liu

2021

Preprint

Self Cite

View full text Add to dashboard Cite

Despite the increasing importance of strain localization modeling (e.g., failure analysis) in computer-aided engineering, there is a lack of effective approaches to consistently modeling related material behaviors across multiple length scales. We aim to address this gap within the framework of deep material networks (DMN) -a physics-based machine learning model with embedded mechanics in the building blocks. A new cell division scheme is proposed to track the scale transition through the network, and its consistency is ensured by the physics of fitting parameters. Essentially, each microscale node in the bottom layer is described by an ellipsoidal cell with its dimensions back-propagated from the macroscale material point. New crack surfaces in the cell are modeled by enriching cohesive layers, and failure algorithms are developed for crack initiation and evolution in the implicit DMN analysis. Besides single material point studies, we apply the multiscale model to concurrent multiscale simulations for the dynamic crush of a particle-reinforced composite tube and various tests on carbon fiber reinforced polymer composites. For the latter, experimental validations on an off-axis tensile test specimen are also provided.

show abstract

Section: Motivation and Backgroundmentioning

confidence: 99%

Section: Appendix a Machine Learning Of Dmn: Data Generation And Opti...mentioning

confidence: 99%

Cell division in deep material networks applied to multiscale strain localization modeling

Liu

2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…This method has also been applied for multi-scale damage modelling where crack paths are predicted for elastoplastic strain softening materials [36]. Liu et al [37] developed a transfer-learning strategy that trains a neural network to predict the stress state of a RVE with a given volume fraction and extend it to any volume fraction. This issue is not limited to computational mechanics, see, e.g., the work of Lu et al [38] on the electrical behaviour of graphene/polymer nanocomposite.…”

Section: Introductionmentioning

confidence: 99%

Machine learning based multiscale calibration of mesoscopic constitutive models for composite materials: application to brain white matter

Field

Ammouche

Peña³

et al. 2021

Comput Mech

View full text Add to dashboard Cite

A modular pipeline for improving the constitutive modelling of composite materials is proposed.The method is leveraged here for the development of subject-specific spatially-varying brain white matter mechanical properties. For this application, white matter microstructural information is extracted from diffusion magnetic resonance imaging (dMRI) scans, and used to generate hundreds of representative volume elements (RVEs) with randomly distributed fibre properties. By automatically running finite element analyses on these RVEs, stress-strain curves corresponding to multiple RVE-specific loading cases are produced. A mesoscopic constitutive model homogenising the RVEs’ behaviour is then calibrated for each RVE, producing a library of calibrated parameters against each set of RVE microstructural characteristics. Finally, a machine learning layer is implemented to predict the constitutive model parameters directly from any new microstructure. The results show that the methodology can predict calibrated mesoscopic material properties with high accuracy. More generally, the overall framework allows for the efficient simulation of the spatially-varying mechanical behaviour of composite materials when experimentally measured location-specific fibre geometrical characteristics are provided.

show abstract

“…To overcome this problem, we previously developed a novel morphology-search method, which is explained in detail below [ 30 ]. To the best of our knowledge, no machine learning (ML) techniques have been applied to analyze complex filler morphologies, although ML was utilized in the material design field called “materials informatics” [ 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 ]. These studies were limited to constructing a microstructure–property linkage using ML, and did not address the mechanisms of the mechanical property.…”

Section: Introductionmentioning

confidence: 99%

Analysis on Microstructure–Property Linkages of Filled Rubber Using Machine Learning and Molecular Dynamics Simulations

et al. 2021

View full text Add to dashboard Cite

A better understanding of the microstructure–property relationship can be achieved by sampling and analyzing a microstructure leading to a desired material property. During the simulation of filled rubber, this approach includes extracting common aggregates from a complex filler morphology consisting of hundreds of filler particles. However, a method for extracting a core structure that determines the rubber mechanical properties has not been established yet. In this study, we analyzed complex filler morphologies that generated extremely high stress using two machine learning techniques. First, filler morphology was quantified by persistent homology and then vectorized using persistence image as the input data. After that, a binary classification model involving logistic regression analysis was developed by training a dataset consisting of the vectorized morphology and stress-based class. The filler aggregates contributing to the desired mechanical properties were extracted based on the trained regression coefficients. Second, a convolutional neural network was employed to establish a classification model by training a dataset containing the imaged filler morphology and class. The aggregates strongly contributing to stress generation were extracted by a kernel. The aggregates extracted by both models were compared, and their shapes and distributions producing high stress levels were discussed. Finally, we confirmed the effects of the extracted aggregates on the mechanical property, namely the validity of the proposed method for extracting stress-contributing fillers, by performing coarse-grained molecular dynamics simulations.

show abstract

Transfer learning of deep material network for seamless structure–property predictions

Cited by 48 publications

References 31 publications

Cell division in deep material networks applied to multiscale strain localization modeling

Cell division in deep material networks applied to multiscale strain localization modeling

Machine learning based multiscale calibration of mesoscopic constitutive models for composite materials: application to brain white matter

Analysis on Microstructure–Property Linkages of Filled Rubber Using Machine Learning and Molecular Dynamics Simulations

Contact Info

Product

Resources

About