Heaviside-1 was trained on 500B field samples from over 250K unique physical layout designs, and is approximately the scale of GPT-2. It predicts a full design's fields ~10^5x faster than a commercial full-wave solver, with S-parameter magnitude error under 1 dB. and generalizes out of distribution.

You can play around with the model and a subset of designs including hairpin filters and patch antennas with our Fields Studio. Design, adjust the geometry, and the predicted electric and magnetic fields will update in milliseconds.