Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials
Authors: Ivan Grega, Ilyes Batatia, Gabor Csanyi, Sri Karlapati, Vikram Deshpande
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we generate a big dataset of structure-property relationships for strut-based lattices. The dataset is made available to the community which can fuel the development of methods anchored in physical principles for the fitting of fourth-order tensors. In addition, we present a higher-order GNN model trained on this dataset. We compare the model to non-equivariant models based on a number of error metrics and demonstrate its benefits in terms of predictive performance and reduced training requirements. ... Table 1: Performance of the different models and training strategies ... Figure 2: The evolution of (a) component loss, Lcomp, (b) equivariance loss, Lequiv, and (c) percentage of negative eigenvalues, λ %, during training. |
| Researcher Affiliation | Collaboration | Ivan Grega1 , Ilyes Batatia1, Gabor Cs anyi1, Sri Karlapati1,2, , Vikram S. Deshpande1 1 Department of Engineering, University of Cambridge; 2 Amazon Science |
| Pseudocode | No | No explicit pseudocode or algorithm block was found within the paper. The methodology is described in narrative form. |
| Open Source Code | Yes | The code is available publicly at github.com/igrega348/energy-equiv-lattice-gnn.git. |
| Open Datasets | Yes | Datasets are available publicly at doi.org/10.17863/CAM.106854. |
| Dataset Splits | Yes | For the machine learning tasks in this paper, we selected from base lattices: (i) 7000 training base lattices, and (ii) 1296 validation/test base lattices. This split ensures that we do not have similar perturbations of the same lattice in both training and test sets. Validation set consists of the 1296 lattices without any perturbations. |
| Hardware Specification | Yes | All models were trained on a single NVIDIA A100 GPU with 80GB of memory. The tests were run on desktop computer with Intel i7-11700 CPU, 96GB RAM and Nvidia RTX3070 GPU. |
| Software Dependencies | No | The training routines were handled by Pytorch Lightning. Optimizer Adam W was used with settings (β1, β2) = (0.9, 0.999), ϵ = 1 10 8, weight decay=1 10 8. (No specific version numbers for PyTorch or Pytorch Lightning were provided.) |
| Experiment Setup | Yes | Hyperparameters were searched on a grid (Table 5). Every experiment was run with constant learning rate for up to 100 000 steps. Optimizer Adam W was used with settings (β1, β2) = (0.9, 0.999), ϵ = 1 10 8, weight decay=1 10 8. Validation loss was checked every 100 steps and training was stopped by early stopping callback with patience 50 if validation loss has not reduced. ... A smaller batch size of 64 was used because of the higher memory requirements of the MACE model. To maintain consistency with the batch size of 256 from CGC models, gradient accumulation over 4 batches was used. |