iph.so
iph.so is building neural systems for physical modeling: fast representations, learned operators, and simulation-aware architectures for video, fluids, materials, and complex dynamical systems.
Neural Physics
Language models learn language. Vision models learn images.
Building the computational foundations for AI that understands the physical world.
Physical systems are different. They evolve through space and time, obey conservation laws, and admit rich mathematical structure. We believe they require their own computational abstractions — not simply larger models or more data.
At iph.so, we call this Neural Physics: the study of representations, operators, and computational systems for learning, predicting, and reasoning about physical dynamics.
Current Research Directions
Each direction explores a different piece of Neural Physics. They are not independent projects — advances in one are expected to transfer across the others. Video representations inform operator learning. Scientific computing shapes new architectures. Browser infrastructure makes the resulting systems practical to inspect, reproduce, and deploy.
Video is our experimental laboratory for Neural Physics. Rather than pursuing compression as an end in itself, we study compact latent representations that preserve the structure needed for prediction, manipulation, and physical reasoning.
Instead of learning observations, we learn the transformations that govern how latent physical states evolve. Our work explores transport operators, conservation-aware architectures, and long-horizon stability for scientific systems.
We study how Neural Physics can accelerate scientific simulation, optimization, and engineering design. Fusion is our first major application area, spanning plasma transport, surrogate simulation, tokamak optimization, and stellarator design — but the underlying ideas extend far beyond any single scientific domain.
Powerful neural systems should be inspectable, reproducible, and broadly accessible — not confined to hyperscale clusters. We build browser-native runtimes, compact model formats, and WebGPU execution systems that make frontier research practical to run and share on commodity hardware.
These are our current working groups, not fixed boundaries. As the field of Neural Physics matures, we expect new directions to emerge while existing ones increasingly converge.
Browse Everything We Track →Fellowship
The iph.so Fellowship exists to remove friction for ambitious independent researchers — support, compute, and a small community working on hard problems together. It isn't employment, and we won't tell you what to work on.
