Shape the geometry.
Hear the response.
Explore how five physical dimensions change the predicted ultrasonic pressure curve—instantly, before committing to a design.
Explore a design.
Forward LightGBM — gradient-boosted trees, geometry → curve Chosen because 353 simulated sweeps is far too little data for a deep network, and boosted trees handle the sharp, non-linear way resonance position and width shift with geometry. Each tree ensemble predicts one shape parameter, not one pressure value.
The model never predicts 360 pressures directly. It predicts the shape of the response — a sloping baseline plus up to three Gaussian resonances — and the curve is rebuilt from those parameters, so every prediction stays a physically plausible resonance.
- Set the geometry. Drag a slider or type an exact millimetre value. Both are clamped to the range the model was trained on.
- Read the response. The curve and the maximum-pressure headline redraw as you move; the three peak cards give centre frequency, amplitude and bandwidth.
- Check the third peak. It only appears when the classifier says it should — its presence confidence is printed on the card. Reset nominal returns you to the reference design.
Stay inside the learned design space.
Controls are limited to the geometry ranges used to train the model.
Measure it backwards.
Inverse 1D CNN — tuned convolutional network, curve → geometry Chosen because the input is a 61-point sweep, not five independent numbers. A convolution shares its filters across frequency, so it learns peak shapes and the spacing between them wherever they sit in the band — which is exactly what identifies a geometry.
This is stage 01 run in reverse: given a response you measured, which five dimensions most likely produced it? Useful for reverse-engineering a sample, or checking whether a manufactured part matches its drawing.
- Load a sweep. Pick a held-out test example from the dropdown, or upload a CSV — frequency in Hz, pressure in Pa — or paste the rows directly.
- Infer the geometry. The CNN returns all five dimensions in millimetres. Test examples also show the known geometry alongside, so you can see the error.
- Set your tolerance. Pick ±1, 2 or 5% and each dimension is marked inside or outside it — the manufacturing question, answered directly.
No sweep loaded yet.
Choose a held-out test example to compare the CNN estimate with its known geometry.