Multi-Agent Language-Guided Metamaterial Discovery via Symbolic Latent Evolution
MetaSymbO turns natural-language design goals into three-dimensional metamaterial lattices. A language-based designer, a learned generator, and a supervisor collaborate through symbolic operations and latent-space optimization to explore candidate structures and their predicted elastic properties.
We appreciate the citations as follows:
@inproceedings{
chen2026metasymbo,
title={MetaSymbO: Multi-Agent Language-Guided Metamaterial Discovery via Symbolic Latent Evolution},
author={Jianpeng Chen and Wangzhi Zhan and Dongqi Fu and Junkai Zhang and Zian Jia and Ling Li and Wei Wang and Dawei Zhou},
booktitle={Third Conference on Language Modeling},
year={2026},
url={https://openreview.net/forum?id=XccQjdb8wh}
}
Check out our demo introduction video: Demo Video →
Explore existing designs or generate a new lattice directly in your browser. No installation, SSH tunnel, account, or personal API key is required.
An actual generated candidate in the live interface. Predictions and supervisor feedback are labeled separately from physical simulation.
| Workflow | What you can do |
|---|---|
| Explore | Browse and search saved lattices; rotate, zoom, inspect nodes, and repeat unit cells in 3D. |
| Generate | Describe a mechanical design goal, review the brief, and follow the asynchronous generation process. |
| Predict | Run the local property predictor and inspect Young's modulus, shear modulus, and Poisson-ratio components. |
| Compare | Inspect up to four candidates with synchronized cameras, consistent display scales, and geometry/property tables. |
| Review evidence | Examine connectivity checks, supervisor scores, suggested refinements, and recorded generation settings. |
| Keep results | Reopen completed runs; import NPZ lattices and export geometry, CSV tables, viewport images, and JSON reports. |
Start here: open the demo → choose Explore Existing Designs or Generate New Design → inspect the structure and its evidence → compare or export your results.
Example brief:
Design a simple connected lattice that is stiffer along Z than X and Y, with fewer than 20 nodes.
The shared demo runs one model job at a time. Generation is limited to two requests per hour per IP, with bounded attempts and runtime. Public runs belong to your browser session; export important results before the seven-day cleanup window. The service currently uses HTTP, so avoid confidential inputs.
Interface guide · Deployment and operations · Service health
- Designer agent translates the design brief into a structural scaffold.
- Generator model uses the VAE, symbolic latent operations, and optimization to construct candidate lattices.
- Supervisor agent evaluates candidates using predicted elastic properties and proposes refinements.
- Design Studio exposes the geometry, predictions, feedback, progress, and saved results for human review.
Supported symbolic operations are union, mix, int, and neg. They operate in the learned representation and do not guarantee exact geometric Boolean operations or satisfaction of every requested constraint.
The deployed application is tested with Python 3.11 in the project's mat_env environment. Model execution requires the project's PyTorch/PyTorch Geometric stack and compatible checkpoints. Browsing additionally uses FastAPI, Uvicorn, NumPy, and Plotly; there is no frontend build step or CDN dependency.
git clone https://github.com/cjpcool/MetaSymbO.git
cd MetaSymbO
conda activate mat_env
python -m WebInterface.app --port 18766Open http://127.0.0.1:18766. This command assumes the existing model environment is installed; cloning the repository does not download the dataset or model weights. Port 18765 is reserved for the deployed backend on zhoulab-1.
For generation, configure OPENAI_API_KEY in a server-side environment variable or the Git-ignored root .env, and restrict the file to its owner:
chmod 600 .envThe server reads the key without executing the file. Credentials never reach browser JavaScript. Prediction does not require an API key; generation also requires the reference dataset. The interface reports which capabilities are ready.
Set METASYMBO_DATASET to the prepared LatticeModulus directory and METASYMBO_CHECKPOINT to a directory containing:
best_ae_model.pt
best_predictor_model.pt
The web worker defaults to CPU, gpt-4o-mini for the designer, and gpt-4.1 for the supervisor. See the configuration table for paths and overrides. Deployment binaries, credentials, checkpoints, and private run data are not published as web assets.
The lattice dataset is handled by datasets/dataset_truss.py. Provide the trusted raw LatticeModulus data under the directory expected by the dataset loader, then prepare its processed data. The web workflow expects data/processed/data.pt beneath METASYMBO_DATASET. The current dataset URL is unset in source; automatic download is not configured.
The research code includes:
| Location | Purpose |
|---|---|
model/ |
Agent collaboration and symbolic latent operations |
modules/ |
Geometric VAE, predictor, and supporting model components |
datasets/ |
Lattice data loading and preprocessing |
train_ae.py, train_predictor.py |
Research training and predictor evaluation scripts; review their local paths and execution settings before use |
evaluation/ |
Generation, prediction, and representation evaluation |
visualization/ |
Lattice visualization and existing voxel/homogenization utilities |
WebInterface/ |
Browser interface, API, isolated model worker, and tests |
deploy/ |
Persistent services, reverse proxy, operations, and public-run cleanup |
Model checkpoints must be supplied separately. The public demo already has its required weights and dataset configured; users of the hosted demo do not need to obtain them.
For a command-line generation run, set OPENAI_API_KEY securely in your environment first, then run:
python run_modal_agent.py \
--prompt "Design a connected lattice with high stiffness along Z and fewer nodes." \
--dataset_path /path/to/LatticeModulus \
--cuda -1 \
--logic_mode union \
--max_collaboration_num 2 \
--save_dir results/latticesUse python run_modal_agent.py --help for the complete options. --condition_vec accepts an optional list of 12 target property values; --verbose enables visualization. The public web demo applies its own conservative resource limits.
The interface distinguishes geometry, model predictions, supervisor feedback, and physical simulation. Stored conditioning vectors are never presented as predictions. Property components retain the model's order; physical units and directional mapping remain unverified. Connectivity checks and supervisor scores do not establish mechanical stability, manufacturability, or validated material performance.
The Validate tab currently records simulation assumptions; it does not execute a solver. For the existing physics-based workflow, see PhyVer. Its deployed pipeline uses UMA optimization and optional ORCA DFT for atomistic materials. That workflow is a simulation reference, not a direct solver for MetaSymbO's mechanical truss lattices. MetaSymbO already contains the same voxelization and homogenization utilities as the PhyVer repository; no second implementation is needed.
Run the automated checks in the model environment:
python -m unittest WebInterface.test_app -v
node --check WebInterface/static/app.jsThe suite covers geometry, supervisor parsing, imports/exports, session isolation, input limits, rate limits, path traversal, timeouts, and credential-safe errors. It makes no paid API requests. Production validation also covers live prediction/generation, desktop/mobile browser behavior, restart persistence, and independent off-host connectivity checks.
The hosted demo uses Caddy on public port 5559 → FastAPI on 127.0.0.1:18765 → an asynchronous model worker. User systemd services keep it running after logout and reboot. See the deployment guide for installation, start/stop/restart commands, logs, retention, and the remaining HTTPS certificate setup.