10104nas a2200133 4500000000100000008004100001260001500042100002100057245011700078856004600195300000700241490012700248520959500375 2026 d c30.10.20261 aMichał Gorawski00aFrom Text to Simulation: Building EPANET Models from Formalized Network Descriptions with a Local Language Model uhttps://www.mdpi.com/2073-4441/18/21/2721 a230 vWater 2026, 18(21), 2721; Special Issue: Looking into the Future of Smart Water Management by Artificial Intelligence (AI)3 a

Hydraulic models of water distribution networks are the basis of a digital twin, but they are still built mostly by hand. HydraKG is a software pipeline that constructs an EPANET model from a network description and checks the result against that description. In the tabular path, a deterministic parser reads GIS-style attribute tables. In the verbal path, a local large language model (LLM) reads a formalized natural-language description. The network is then checked by validators, by the EPANET solver and by a comparison with the description, none of which uses the source model; code repairs every fact with a single correct value, and the LLM corrects the rest in up to three rounds before the final model is simulated. On seven networks with 11 to 785 nodes, the tabular path reproduces the junction pressures of the source models, as represented internally, with a root-mean-square error (RMSE) of at most  m, which is the rounding limit. On canonical descriptions, the verbal path gave a model equal in value to the description in all 52 runs on six networks, including a real distribution zone. On the larger networks, the first extraction omitted groups of elements in every run, and correction rounds restored them.