Training neural networks to simulate physics for complex hardware like antenna arrays just got dramatically cheaper.
A paper titled "PE-EK-PINN: Physics Embedding with Evolving Kernel for Scalable Physics-Informed Neural Networks" (arXiv:2609.38023) describes the fix. Physics-informed neural networks (PINNs) are meant to learn physical laws directly, but most only enforce those laws as a soft penalty in the loss function, leaving highly oscillatory wave behavior for the optimizer to stumble into. That works on textbook benchmarks but breaks down on real radiation problems with singular excitations and absorbing boundaries. An earlier fix hard-coded physics into the architecture using a dictionary of oscillatory kernels, but that dictionary has to be built by hand and grows exponentially for hierarchical systems, including antenna arrays and metasurfaces built from many repeated elements.
PE-EK-PINN's trick is to treat a converged solution for one subsystem as a reusable kernel, then transform and reuse it to assemble larger systems instead of deriving new equations for every added unit. That keeps the number of active kernels flat regardless of system size and cuts cumulative training cost from scaling linearly with the number of units to scaling with its logarithm. On a 256-dipole array, the paper reports training more than 30 times faster than a direct physics-embedded PINN, with accuracy that matches or beats the older method.
The appeal isn't a new physics discovery. It's an engineering shortcut that could make PINNs practical for radar, wireless, and metasurface design at scales that were previously too slow to simulate this way. It also tracks a broader shift in scientific machine learning toward reusing learned representations instead of just throwing more compute at bigger models.
The 30x figure comes from a single 256-dipole benchmark, not a general guarantee, so how well the trick holds up on messier, real-world hardware is still an open question.