TL;DR
This paper introduces a novel, publicly available deep-learning framework that bridges PyTorch and ABAQUS for modeling plasticity in materials. It allows engineers to integrate complex, data-driven plasticity models into their finite element analyses without needing to write extensive FORTRAN code, significantly reducing computational cost and improving accessibility.
What Problem Does This Paper Address?
Machine learning has shown great promise in developing constitutive models for materials, but integrating these models into established engineering simulation software like ABAQUS has been a significant hurdle. Traditional approaches often require complex FORTRAN programming for user-defined material subroutines (UMAT/VUMAT), and the "black box" nature of neural networks makes validating and reproducing results challenging. This paper addresses the need for a practical, transparent, and accessible framework that allows engineers to leverage deep learning for plasticity modeling without deep expertise in both machine learning and FORTRAN programming.
Methods
The core of this work is a framework that separates the training of neural networks (using PyTorch) from their deployment in ABAQUS. Key methodological contributions include:
- Level-Set Plasticity: The yield surface is implicitly represented by a scalar-valued function, allowing for more flexible and complex yield criteria.
- UMAT/VUMAT Integration: A workflow is established to convert trained PyTorch neural networks into generic FORTRAN code that can be directly used as ABAQUS UMAT/VUMAT subroutines. This bypasses the need for manual FORTRAN coding of the neural network architecture.
- Split Training: The learning task is split into two parts: predicting the yield surface evolution and predicting the plastic flow direction. This approach avoids computationally expensive backpropagation during prediction.
- Sobolev Training: An H¹-loss function is employed for training the yield function neural network, which incorporates higher-order derivative information to improve prediction accuracy and smoothness.
- Separate Plastic Flow Network: An additional neural network is introduced to directly predict the stress gradient, further reducing computational cost compared to relying solely on backpropagation.
- Data Augmentation: A level-set approach is used to augment the training data, creating synthetic stress points to help the neural network learn the implicit representation of the yield surface.
- Reproducibility Focus: The authors emphasize making all data, source code, and trained models publicly available to facilitate third-party validation and reproducibility.
Datasets
The paper utilizes datasets generated from benchmark plasticity models, specifically the Tresca, Hosford (nн=1.5), and von Mises yield criteria. These datasets are augmented using the level-set approach to create training data for the neural networks. For the anisotropic extension, data is generated based on the Yld2004-18p criterion.
Key Insights and Ideas
- Bridging the Gap: The framework effectively bridges the gap between the Python-based machine learning ecosystem and the FORTRAN-based finite element analysis world, making advanced constitutive models more accessible to engineers.
- Computational Cost Reduction: Splitting the prediction of the yield function and its gradient into two separate neural networks significantly reduces the computational cost during the stress integration phase, by orders of magnitude compared to previous methods.
- Sobolev Training Benefits: Using H¹-loss for yield function training demonstrably improves prediction accuracy and smoothness, leading to better model performance.
- Weak Interrelation of Tasks: Comparing single-task and multi-task neural networks suggests that predicting the yield function and its stress gradient are weakly interrelated tasks, implying that separate networks might be more effective.
- Importance of Data Pre-processing: The study highlights the critical role of data denoising in preventing overfitting and achieving satisfactory training goals, especially for complex yield surfaces.
Key Results
- The framework successfully integrates trained neural networks into ABAQUS UMAT/VUMAT subroutines.
- Numerical experiments show that the trained models can accurately recover the yield surface and hardening behavior of benchmark plasticity models (Tresca, Hosford, von Mises).
- The proposed approach using separate neural networks for yield function and stress gradient prediction is shown to be significantly faster (two orders of magnitude) than methods relying on backpropagation for gradient calculation.
- The framework is extended to capture anisotropic plasticity behavior, demonstrating its versatility.
- The study shows that Sobolev training (H¹-loss) leads to lower prediction errors compared to L²-loss for yield function prediction.
Limitations and Future Work
- Limited Neuron Types: The framework currently supports feed-forward neurons, and extensions to more elaborated neuron types are considered future work.
- Data Pre-processing: While the paper demonstrates the benefit of data denoising, the source code for the denoising technique itself is not included, though scripts for denoising are provided.
- Specialized Architectures: The paper mentions the potential of specialized neural network architectures (e.g., input convex neural networks) to enforce convexity, which is outside the scope of this study but a promising future direction.
- Hyperparameter Tuning: The neural network architectures used were not extensively tuned, suggesting that further optimization could lead to even better performance.
- Extension to Other Models: While the framework is demonstrated for plasticity, its application to other material models and complex phenomena is a natural extension.
Summary updated