GranumFL: an agentic, GPU-accelerated coupled DEM–SPH program for debris flows on real terrain

Debris flows on real hillsides involve grains, water and terrain together, yet the programsthat couple discrete elements with smoothed particle hydrodynamics on a GPU are eithercommercial and resolved at the grain scale or research codes that an engineer cannot setup without programming. This paper presents GranumFL, which couples a weakly com-pressible SPH water phase with a DEM grain phase through the unresolved formulation ofRobinson, Ramaioli and Luding (2014), runs both phases on one CUDA GPU, lays out acase on a digital elevation model, and exposes every step to a large-language-model agentthrough 20 Model Context Protocol tools. Each phase was calibrated against publishedbenchmark results. The DEM phase reproduces the Chung and Ooi impact benchmarkswithin 0.09 % of Hertz’s solutions and 0.19◦ of Kratos, and the DEM8 penetration testwithin the spread of eight open-source codes, in 1.1 min on an RTX 5090, as fast as thededicated GPU codes. For a sedimenting porous block, the coupled phase is closer totheory than the reference implementation of Robinson et al. at every resolution compared(−2.7, +9.3 and +14.4 % against +5, +15 and +22 % at h/d = 2, 5 and 6), whereas for asingle grain it is accurate only for 2.5 ≤ h/d ≤ 4, because of the grain’s own volume in theporosity at small h/d and a truncated buoyancy at large h/d, in a dense fixed bed it gives1.5 to 3.3 times Ergun’s pressure drop, and a loose submerged column collapses too farwith a base pore pressure of the wrong sign, so dense granular assemblies with their porepressure remain the open problem. The SPH phase follows the Martin and Moyce dambreak within 3.0–3.1 % RMS and hydrostatics within 0.68 %, and in SPHERIC Test 2 itfollows the water heights within about 8 % and times the impact on the box within 3 msbut over-predicts the peak pressure near the floor, at about the speed of DualSPHysicson a V100 at the coarser spacing. The CPU and GPU agree to three digits, and the waterphase sustains 1.6–1.9×108 particle updates per second up to 12 million particles, 55 to 66times a 30-thread CPU. In four independent trials an LLM agent, driving only the MCPtools, set up and edited the 2026 Patong (Phuket) debris-flow site from plain-languagerequests and passed all 33 objective checks, and two trials found on their own that thesteepest-descent path on the Copernicus surface model does not follow the reported route.