%0 Conference Paper %1 sommer2021automation %A Sommer, Daniel %A Vinod Kumar Mitruka, Tarun Kumar Mitruka %A Middendorf, Peter %B Proceedings of the 13th European LS-DYNA Conference 2021 %D 2021 %I DYNAmore GmbH %K from:maltevonscheven %T Automation of LS-DYNA’s Material Model Driver for Generation of Training Data for Machine Learning based Material Models %U https://www.dynalook.com/conferences/13th-european-ls-dyna-conference-2021/material-modeling/sommer_university_of_stuttgart.pdf/@@download/file/Sommer_University_of_Stuttgart.pdf %X The substitution of classical constitutive material models with data-driven models supported by machine learning techniques could provide a leap in the modelling of materials. The most notable benefits are a faster description of new materials without a tedious manual parameter identification procedure, lower computational time for simulations due to efficient computation within the material model and a more efficient selection of the correct material model for the use-case. The base for any data-driven model is adequate amount and quality of training data. Based on this, machine learning techniques can be used to train neural networks such that they learn the relationship between given input and output. The mapping in the machine learning based material model will be the strain measures to the stresses, similar to classical models. In order to learn the stress-strain relationship for materials, training data is generated from existing material models in LS-DYNA, as the direct extraction from materials testing is not possible due to the impossibility of local stress measurements. To generate training data, an existing model could be implemented in own code, single-element simulations could be performed, or even data from component simulations could be extracted. This study shows a way to generate training data for almost any material model available in LS-DYNA by using and automating the integrated Interactive Material Model Driver. This enables access to the unbiased stress response of the black-box material models, by providing an arbitrary progression of components of the displacement gradient in time as input, to be evaluated at a single integration point. Automation is added on top of the embedded Material Model Driver regarding the generation of strain paths for 2D and 3D cases. It is followed by automatic execution, data extraction and preparation of data for machine learning. Advantages and disadvantages over other strategies for training data generation will be highlighted, compared and finally an exemplary material model using neural networks trained on data generated with the Automated Material Model Driver is shown. The goal is that networks learnt through such a training data can also be used as a surrogate to develop any new material model.

@inproceedings{sommer2021automation, abstract = {The substitution of classical constitutive material models with data-driven models supported by machine learning techniques could provide a leap in the modelling of materials. The most notable benefits are a faster description of new materials without a tedious manual parameter identification procedure, lower computational time for simulations due to efficient computation within the material model and a more efficient selection of the correct material model for the use-case. The base for any data-driven model is adequate amount and quality of training data. Based on this, machine learning techniques can be used to train neural networks such that they learn the relationship between given input and output. The mapping in the machine learning based material model will be the strain measures to the stresses, similar to classical models. In order to learn the stress-strain relationship for materials, training data is generated from existing material models in LS-DYNA, as the direct extraction from materials testing is not possible due to the impossibility of local stress measurements. To generate training data, an existing model could be implemented in own code, single-element simulations could be performed, or even data from component simulations could be extracted. This study shows a way to generate training data for almost any material model available in LS-DYNA by using and automating the integrated Interactive Material Model Driver. This enables access to the unbiased stress response of the black-box material models, by providing an arbitrary progression of components of the displacement gradient in time as input, to be evaluated at a single integration point. Automation is added on top of the embedded Material Model Driver regarding the generation of strain paths for 2D and 3D cases. It is followed by automatic execution, data extraction and preparation of data for machine learning. Advantages and disadvantages over other strategies for training data generation will be highlighted, compared and finally an exemplary material model using neural networks trained on data generated with the Automated Material Model Driver is shown. The goal is that networks learnt through such a training data can also be used as a surrogate to develop any new material model.}, added-at = {2022-04-20T12:46:04.000+0200}, author = {Sommer, Daniel and Vinod Kumar Mitruka, Tarun Kumar Mitruka and Middendorf, Peter}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/2c14434f6768e3ec2e2024e4469002ec6/ibb-publication}, booktitle = {Proceedings of the 13th European LS-DYNA Conference 2021}, interhash = {3533b7de3cb2f6188449e73182a30db6}, intrahash = {c14434f6768e3ec2e2024e4469002ec6}, keywords = {from:maltevonscheven}, publisher = {DYNAmore GmbH}, timestamp = {2022-04-20T10:46:04.000+0200}, title = {Automation of LS-DYNA’s Material Model Driver for Generation of Training Data for Machine Learning based Material Models}, url = {https://www.dynalook.com/conferences/13th-european-ls-dyna-conference-2021/material-modeling/sommer_university_of_stuttgart.pdf/@@download/file/Sommer_University_of_Stuttgart.pdf}, year = 2021 }