Paper

This repository is the official implementation of Inverse design of mechanical metamaterials balancing manufacturability and compactness: A case study on lattice cells.

Abstract

Mechanical metamaterials are artificially engineered structures designed to exhibit unique and extraordinary mechanical properties. In recent years, machine learning provides a more efficient and systematic approach enabling inverse design of mechanical metamaterials, which allow for a broader exploration of material properties and support the integration of multifunctionality, significantly speeding up the design process. Despite the many advantages of inverse design, metamaterials often involves a trade-off between competing performance metrics-such as manufacturability and structural compactness. Furthermore, these trade-offs should be dynamically adjusted based on different additive manufacturing conditions. To address this, we proposed a regressional and conditional generative adversarial network based multi-objective (RCGAN-MO) architecture, which simultaneously handles the inverse design and adjustable multi-objective optimization of mechanical metamaterials. The RCGAN-MO includes two trained neural networks: a generator and a predictor, along with a weighted multi-objective optimizer. As a case study, the RCGAN-MO architecture is applied to the inverse design of the relative compressive elastic modulus for a metamaterial, and metamaterials with different weight vector values in the multi-objective optimizer are achieved through 3D printed prototypes. This approach achieves high accuracy and could adjust the importance of manufacturability and compactness, offering a flexible, scalable solution for engineering metamaterials tailored to practical application demands.

File Structure

├── ML/
│   ├── configs/nn.json                     # NN configuration
│   ├── dataset/                            # Training & validation data
│   │   ├── data_for_ml.csv
│   │   └── FEM_val_data.csv
│   └── models/
│       ├── train_config.py                 # Model configuration
│       ├── nn/                             # Forward neural networks
│       │   ├── forward_train.py
│       │   └── forward_eval.py
│       └── gan/                            # RCGAN implementation
│           ├── rcgan_train.py              # GAN training
│           ├── rcgan_eval.py               # GAN evaluation
│           ├── multi-objective_optimizer/  # Multi-objective optimization
│           │   └── gradient_optimization.py
│           └── plot_results/               # Visualization tools
├── Exp_results/output_moduli.csv           # Experimental results
├── figs/gen_img.jpg                        # Generated images
├── PTH_files/                              # Trained models
└── requirements.txt                        # Dependencies

Usage

1. Training: Run rcgan_train.py for RCGAN model training
2. Optimization: Use gradient_optimization.py for multi-objective optimization
3. Evaluation: Execute rcgan_eval.py for model evaluation and visualization