← SPECTRAL / SHAPE LAB SIMPLE

Optimize from a saved model.

Choose vector10_relu or vector10_silu. The app evaluates and differentiates the saved network entirely in your browser. It uses no training samples, stored optimizers or nearest-neighbor support restriction.

Run and save

  1. Choose a model and a functional using x1 through x10.
  2. Choose area, direction, number of starts, step budget and random seed.
  3. Start the search. The first start is the disk; the rest are generated random feasible shapes, without ranking them by predicted quality.
  4. Pause, resume or stop. Save JSON, Fourier coefficients CSV, SVG or PNG during or after a run.

The default is 16 starts and 300 steps. A single start uses only the disk. For the same seed and bounds, both models receive the same starting shapes. Changing the model clears the previous result.

Write a functional

x5
x6
x5 + 0.5*x6
x6/x1 + 0.01*log(x10)

Use arithmetic, parentheses, powers, sin, cos, tan, asin, acos, atan, sinh, cosh, tanh, exp, log/ln, sqrt, pi and e. Write multiplication explicitly. Variables are physical eigenvalues at the requested area. Formula derivatives are symbolic and combined with neural derivatives. Undefined trial values trigger backtracking. The formula must be defined at the disk.

Geometry

The area is normalized analytically. Projected amplitude bounds keep r(θ) ≥ 0.25 a₀ > 0 for the supplied models, so boundaries remain simple. Harmonic caps stay active. Spectra and objectives are predictions, with no PDE solve. A stopped search is not a proof of optimality.

Bundle from .pt

Put one checkpoint path per line in web_simple/model-paths.txt and run:

python web_simple/scripts/update_app.py

Or bundle a specific file:

python web_simple/scripts/update_app.py --pt path/to/my_model.pt

The offline converter reads weights and normalization from .pt and checks the browser's predictions and gradients. It does not need a dataset or PDE software. The deployed browser loads the generated compact bundle; it does not need Python. See the checkpoint contract and technical instructions.

Serve or publish

python3 -m http.server 8000 --bind 127.0.0.1 --directory web_simple

For GitHub Pages, publish the contents of web_simple/dist/, which the update script prepares. Any static host works, including a project subdirectory. No computation backend is needed.

Return to the app →