KAFlow preprint reports more accurate learning of physical dynamics

Researchers report in an arXiv preprint that KAFlow, a neural network method for learning physical dynamics, lowers loss on data outside its training distribution by 3.4 to 11 times on synthetic benchmarks compared with the Kolmogorov-Arnold neural operator. It also recovers every coefficient to within 10^-4 without symbolic templates. KAFlow represents the mathematical rule governing the dynamics through a small set of single-variable functions and automatically extracts a closed-form expression. On long-horizon quantum dynamics, it reduces state infidelity by up to 6 orders of magnitude, with a 30 times reduction when trained only on position and momentum measurements. These findings come from benchmark tests.