DEM + SPH · one CUDA GPU · agentic
Grains and water, coupled on one GPU. GranumFL adds an SPH water phase to the Granum discrete element program, lays out debris flows on real terrain from a GeoTIFF, and lets an AI agent set up and read a site study through 20 MCP tools.
A 2D toy running in your browser (grains coloured by speed, water in blue). GranumFL itself is 3D WCSPH–DEM on the GPU.
What it does
Depth-averaged models replace boulders and water by one continuum. GranumFL keeps the grains as discrete elements and the water as SPH particles, so the boulders that dam a channel, the pore water that drains from a deposit and the erosion of a bed all come out of the physics.
The locally averaged formulation of Robinson, Ramaioli and Luding (2014): grains smaller than the smoothing length, a porosity field, Di Felice drag and −V∇P on each grain, with the exchanged momentum balanced to round-off.
CUDA kernels compiled at run time with NVRTC (sm_120 included). The water reads the grains from the DEM buffers on the device, so a coupled step moves no particle data to the host. The same algorithm runs on the CPU in double precision.
Open a GeoTIFF or ESRI grid, or download the Copernicus GLO-30 tile of a latitude and longitude, crop, resample and cut. WGS 84 and UTM read from the file.
Debris thickness, erosion and deposit, peak speed, flood depth by volume, runout and travel angle, and records at named places (thickness, speed, water depth, ρv²), as PNG maps or ESRI grids for GIS.
A Model Context Protocol server built into the program. The in-app agent or any MCP client sets up the site, checks the warnings, runs or exports the case and reads the results.
Export a case for a CUDA server, run it with one command and load its results folder back into the app. A million grains over 150 s of flow ran in 63 min on one RTX 5090.
Architecture
An engineer works in the desktop app or asks the agent; both reach the site layer, which lays out the case for the coupled solver. Large cases go to a GPU server and come back as a results folder.
Real terrain
Copernicus GLO-30 is a surface model: canopy and buildings are in its heights, so a lidar terrain model gives better flow paths where one exists.
Agentic
In four independent trials an LLM agent, given only the engineer's request and the MCP tools, built or edited the 2026 Patong (Phuket) debris-flow site and passed all 33 objective checks, scored from the models it left. Two trials found on their own that the steepest-descent path on the surface model misses Nanai Road. Below is a replay of a real trial (T4, five tool calls).
Calibration
Each phase was compared with published results of established codes on the same problems, in the geometry of the original studies. No other code was run here, so the comparisons rest on their published numbers.
| Phase and test | Reference | GranumFL | Reference result |
|---|---|---|---|
| DEM · Chung & Ooi impacts 1–2 | Hertz, Kratos | force and overlap within 0.09 % | exact (analytic) |
| DEM · Chung & Ooi oblique impact | Kratos | rebound angle mean 0.19° above 15° | – |
| DEM · DEM8 penetration, 25,000 grains | 8 open codes (Dosta et al. 2024) | rebound −0.044 m in 1.1 min (RTX 5090) | −0.048 to −0.024 m; 1.2–122 min |
| Coupled · sedimenting porous block, ε = 0.8, h/d = 2 / 5 / 6 | MercuryDPM (Robinson et al. 2014) | −2.7 / +9.3 / +14.4 % | +5 / +15 / +22 % |
| Coupled · porous block, ε = 0.6–0.9 | Robinson et al., Eq. (36) | −2.7 to +5.0 % | within 5 % |
| SPH · dam break, three resolutions | Martin & Moyce (1952) | 3.0–3.1 % RMS front error | – |
| SPH · hydrostatic tank | ρ₀ g (H − z) | 0.68 % (GPU), 1.05 % (CPU) | – |
GPU performance
The water phase sustains 1.6 to 1.9×10⁸ particle updates per second from 0.2 to 12.2 million particles on an RTX 5090, 55 to 66 times 30 CPU threads of an EPYC 7502. The coupled phase advances 3.5 to 4.8×10⁷ grain steps per second up to 4.6 million grains.
For scale only: the DualSPHysics run of Crespo et al. (2015) on a GTX Titan corresponds to at least 2.9×10⁷ particle updates per second; that was older hardware and a mature, optimised code.
Gallery
Download
The same coupled solver in both editions. The Education edition is free for teaching and non-commercial academic research and limits the size of a model, so it runs the bundled templates and small site studies on your own computer.
For students and teachers
Free
For teaching and non-commercial academic research. Need larger sites for your research? Ask for an Academic Research licence.
For engineering practice
On request
Tell us about your project and we will reply with a licence offer.
Paper
Sompote Youwai, King Mongkut's University of Technology Thonburi, 2026. The paper describes the unresolved DEM–SPH coupling on one CUDA GPU, the calibration of each phase against published benchmarks (Chung and Ooi impacts within 0.09 % of Hertz, DEM8 penetration within the spread of eight open-source codes, a sedimenting porous block closer to theory than the reference implementation), the cases where the coupling is not yet accurate, the GPU performance, and the agent trials that set up a site study through 20 MCP tools.
@misc{youwai2026granumfl,
author = {Youwai, Sompote},
title = {GranumFL: an agentic, GPU-accelerated
coupled DEM--SPH program for debris
flows on real terrain},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23858571},
url = {https://zenodo.org/records/23858571}
}
Contact
AI Research Group, Department of Civil Engineering, King Mongkut's University of Technology Thonburi (KMUTT), Bangkok.
GranumFL is built on Granum, our discrete element program for soil.