Arena Physica

Arena Physica has released Heaviside-1, the second generation of its foundation model for electromagnetism. The model takes arbitrary 3D geometries, materials, and excitations as input and returns full electromagnetic near-field predictions in a fraction of a second, replacing the slow mesh-and-solve loop that traditional EM solvers use to get to convergence. It is aimed at the engineers designing chips, packaging, antennas, radar, and other RF hardware, where simulation time has long been the bottleneck on how many designs a team can actually explore. The launch matters because it reframes what an EM neural surrogate is expected to do. Heaviside-0, released in March, focused on S-parameter predictions, delivering them orders of magnitude faster than traditional solvers with error well below 1 dB. Heaviside-1 moves from those scalar summaries to predicting the full 3D electric and magnetic fields, which are the primitives from which downstream quantities like energy flow can be computed. According to the launch, the model was trained on roughly 50 years of accumulated simulation data spanning about a quarter million geometries and hundreds of billions of field labels, and its redesigned encoder can ingest structures as intricate as ball grid arrays in semiconductor packaging. Arena Physica is also promising an EM benchmark suite for this new class of model in the coming weeks. The company is based in New York and was founded by Pratap Ranade and Engin Ural , with Michael Frei serving as Chief Research Officer and presenting the launch. Prospective users can poke at Heaviside-1's field outputs through Atlas Field Studio at studio.arenaphysica.com, which offers a visual entry point into the model before deeper integration.