Sovereign 8s.is mirror of standardgalactic/AI-symmetry-topology-woven-acoustics
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Thanasak Wanglomklang dd33f733db
Revise overview and workflow stages in README
Updated the overview section to reflect changes in the framework description and workflow stages for the inverse design of woven acoustic materials.
2026-06-28 17:00:27 +09:00
DATA Add files via upload 2026-02-21 19:24:26 +01:00
data_preprocessing.py Add files via upload 2026-02-21 19:22:05 +01:00
gpr_training.py Update model path for SemiConvAE 2026-02-22 10:42:30 +01:00
optimization.py Refactor optimization.py: update model paths and remove PCA 2026-02-22 10:41:59 +01:00
README.md Revise overview and workflow stages in README 2026-06-28 17:00:27 +09:00
requirements.txt Add files via upload 2026-02-21 19:22:05 +01:00
semi_convae.py Update semi_convae.py 2026-02-22 10:44:11 +01:00

AI-Assisted Symmetry-Informed Topology Optimization of Woven Materials for Broadband Sound Absorption

Thanasak Wanglomklang, Frédéric Gillot, Sébastien Besset, Sonia Mahmoudi


Overview

This repository provides the code and data accompanying the paper. We propose a symmetry-informed, surrogate-based framework for the inverse design of woven acoustic materials. The workflow follows four stages:

  1. Dataset preparation : generate feasible binary woven patterns under wallpaper symmetries (p2, p4, pm, pmm), hang-on manufacturability constraints, and topological equivalence rules. FEM simulations provide acoustic labels for a subset of patterns, while a larger feasible set is used as unlabeled data.

  2. Semi-ConvAE training : train a semi-supervised convolutional autoencoder to embed 10×10 binary topologies into a compact latent space. An auxiliary regression head is used only during training to organize the latent space with respect to acoustic performance.

  3. GPR surrogate modeling : encode the labeled patterns and train a Gaussian Process Regression surrogate in the learned latent space to predict broadband sound absorption and model uncertainty.

  4. Optimization and reconstruction : optimize the latent vector using a Genetic Algorithm guided by the GPR surrogate, then decode and repair the optimized latent point into a symmetry-valid binary woven topology.

propose_framework

Repository Structure

├── data_preprocessing.py      # Train/test split and labeled subset generation
├── semi_convae.py             # Semi-supervised ConvAE model definition and training
├── gpr_training.py            # GPR surrogate training on the latent space
├── optimization.py            # GA topology optimization + uncertainty quantification
├── requirements.txt           # Python dependencies
└── DATA/
    ├── P2/
    │   ├── labeled_results_objectives.csv   # 300 labeled samples (binary patterns + JCA parameters + alpha)
    │   └── unlabeled_10000.csv              # 10,000 unlabeled feasible patterns
    ├── P4/  (same structure)
    ├── PM/  (same structure)
    └── PMM/ (same structure)

Data Description

labeled_results_objectives.csv (300 rows per symmetry group)

Column group Columns Description
Binary pattern P00 – P99 10×10 binary woven topology (0 = warp over weft, 1 = weft over warp)
Transport parameters porosity, tortuosity, thermal_length, viscous_length, permeability JCA macroscopic parameters from COMSOL FEM
Acoustic output alpha_avg, alpha_max Broadband-averaged and peak sound absorption coefficient

unlabeled_10000.csv (10,000 rows per symmetry group)

Columns P00–P99 only. These are feasible woven patterns satisfying symmetry and hang-on constraints, sub-sampled from a full pool of 1,000,000 patterns and used for unsupervised reconstruction training.


Installation

# Clone the repository
git clone https://github.com/twanglom/AI-symmetry-topology-woven-acoustics.git
cd symmetry-topology-woven-acoustics

# Install dependencies
pip install -r requirements.txt

# Install PyEGRO (GPR library, developed by the authors)
pip install PyEGRO

Usage

Run the scripts in the following order. Before running each script, set the SYMMETRY variable at the top of the file to one of "P2", "P4", "PM", "PMM".

Step 1 — Data preprocessing

python data_preprocessing.py

Splits the 300 labeled samples into a held-out test set (30 samples) and training subsets at six fractions (10%, 20%, 30%, 50%, 80%, 100%). Outputs are saved to DATA/{SYMMETRY}/output/.

Step 2 — Train the Semi-ConvAE

python semi_convae.py

Trains the semi-supervised convolutional autoencoder. Key hyperparameters are set at the top of the main() function:

Parameter Value Description
bottleneck 10 Latent space dimensionality
dropout 0.2 Dropout rate
epochs 5000 Training epochs
reg_weight 10.0 Weight for supervised prediction loss (λ_p)
unsup_weight 3.0 Weight for unlabeled reconstruction loss (λ_u)

Trained model, config, scalers, and latent vectors are saved to DATA/{SYMMETRY}/output/SemiConvAE/.

Step 3 — Train the GPR surrogate

python gpr_training.py

Encodes the training data into the latent space and fits a GPR model with an RBF kernel using PyEGRO. Results are saved to DATA/{SYMMETRY}/output/GPR/.

Step 4 — Topology optimization

python optimization.py

Runs the GA in the latent space to maximize broadband absorption. Performs 30 independent runs for robustness quantification. Outputs include:

  • best_pattern_binary.npy — optimal 10×10 binary woven pattern
  • optimization_report.txt — mean ± std across runs and GPR uncertainty
  • Convergence plots and topology evolution visualizations

Computational Resources

Training was performed on the Newton HPC cluster at École Centrale de Lyon using an NVIDIA Quadro P4000 GPU.


Requirements

torch>=2.0
numpy
pandas
scikit-learn
matplotlib
pymoo
PyEGRO   # pip install PyEGRO  (https://twanglom.github.io/PyEGRO/)

Full list: see requirements.txt.


Full Dataset

The complete dataset is available from the corresponding author upon reasonable request due to file size. COMSOL .mph simulation files are also available upon request.



License

MIT License. See LICENSE for details.


Contact

Thanasak Wanglomklang — thanasak.wanglomklang@ec-lyon.fr
Laboratoire de Tribologie et Dynamique des Systèmes (LTDS), École Centrale de Lyon