A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Deep transfer learning-aided constitutive modelling of granular soils considering out-of-range particle morphology
https://orcid.org/0000-0003-0260-8776
Highlights A large discrete element method simulation database was generated. A data-driven constitutive model of granular soils was established. Long short-term memory-based constitutive model has poor extrapolation ability. Deep transfer learning is adopted to transfer the pre-trained model to the out-of-range data.
Abstract Based on a large discrete element method (DEM) simulation database, a novel data-driven constitutive model of granular soils was established using the long short-term memory (LSTM) neural network. The excellent training and testing results reflect the superiority of LSTM in the constitutive modelling of granular soils under loading. Based on the previous study, twelve new samples with out-of-range particle morphology were generated to test the model extrapolation ability. It is found that the LSTM-based model has poor extrapolation ability with notable predictive differences when out-of-range input parameters are utilised. In light of this deficiency, deep transfer learning (DTL) is adopted to transfer the pre-trained model to the new out-of-range database quickly. The excellent training and testing results show that the DTL-LSTM can accurately predict cases with out-of-range particle morphology. DTL-LSTM is further compared with a newly trained LSTM model using the extended database, and an overall good agreement is obtained except for some slight differences of volumetric strain at small axial strains. Given the drastically reduced computational costs and the excellent predictive performance, DTL-LSTM is deemed to be suitable and competent in the constitutive modelling of granular soils with out-of-range data.
Deep transfer learning-aided constitutive modelling of granular soils considering out-of-range particle morphology
https://orcid.org/0000-0003-0260-8776
Highlights A large discrete element method simulation database was generated. A data-driven constitutive model of granular soils was established. Long short-term memory-based constitutive model has poor extrapolation ability. Deep transfer learning is adopted to transfer the pre-trained model to the out-of-range data.
Abstract Based on a large discrete element method (DEM) simulation database, a novel data-driven constitutive model of granular soils was established using the long short-term memory (LSTM) neural network. The excellent training and testing results reflect the superiority of LSTM in the constitutive modelling of granular soils under loading. Based on the previous study, twelve new samples with out-of-range particle morphology were generated to test the model extrapolation ability. It is found that the LSTM-based model has poor extrapolation ability with notable predictive differences when out-of-range input parameters are utilised. In light of this deficiency, deep transfer learning (DTL) is adopted to transfer the pre-trained model to the new out-of-range database quickly. The excellent training and testing results show that the DTL-LSTM can accurately predict cases with out-of-range particle morphology. DTL-LSTM is further compared with a newly trained LSTM model using the extended database, and an overall good agreement is obtained except for some slight differences of volumetric strain at small axial strains. Given the drastically reduced computational costs and the excellent predictive performance, DTL-LSTM is deemed to be suitable and competent in the constitutive modelling of granular soils with out-of-range data.
Deep transfer learning-aided constitutive modelling of granular soils considering out-of-range particle morphology
https://orcid.org/0000-0003-0260-8776
Highlights A large discrete element method simulation database was generated. A data-driven constitutive model of granular soils was established. Long short-term memory-based constitutive model has poor extrapolation ability. Deep transfer learning is adopted to transfer the pre-trained model to the out-of-range data.
Abstract Based on a large discrete element method (DEM) simulation database, a novel data-driven constitutive model of granular soils was established using the long short-term memory (LSTM) neural network. The excellent training and testing results reflect the superiority of LSTM in the constitutive modelling of granular soils under loading. Based on the previous study, twelve new samples with out-of-range particle morphology were generated to test the model extrapolation ability. It is found that the LSTM-based model has poor extrapolation ability with notable predictive differences when out-of-range input parameters are utilised. In light of this deficiency, deep transfer learning (DTL) is adopted to transfer the pre-trained model to the new out-of-range database quickly. The excellent training and testing results show that the DTL-LSTM can accurately predict cases with out-of-range particle morphology. DTL-LSTM is further compared with a newly trained LSTM model using the extended database, and an overall good agreement is obtained except for some slight differences of volumetric strain at small axial strains. Given the drastically reduced computational costs and the excellent predictive performance, DTL-LSTM is deemed to be suitable and competent in the constitutive modelling of granular soils with out-of-range data.
Deep transfer learning-aided constitutive modelling of granular soils considering out-of-range particle morphology
Xiong, Wei (author) / Wang, Jianfeng (author)
2023-12-05
Article (Journal)
Electronic Resource
English
Particle breakage of granular soils: changing critical state line and constitutive modelling
| Springer Verlag | 2022
Particle breakage of granular soils: changing critical state line and constitutive modelling
| Springer Verlag | 2022