TellusPM follows soil from the first plastic strain to where the debris stops: footings, slope failures, excavations and debris flows on real terrain, in 2D or 3D, on a multi-core CPU or an NVIDIA GPU.
points in one whole-hillside debris-flow run, 12 min on one RTX 4090
the CUDA engine against 15 CPU threads, with the same results
typed MCP tools in engineering units, for an engineer or an AI agent
in the verification and calibration campaign of the paper
The method
A finite element mesh tangles long before a landslide stops. In the material point method the soil is a cloud of points that keep their mass, velocity, stress and history, while a fixed background grid solves the momentum balance each step and is then reset, so the points can travel any distance.
Mass, momentum and internal force go to the nodes through uGIMP, linear or quadratic B-spline weights.
Nodal momentum is updated with gravity, local damping, side conditions and frictional contact with rigid bodies.
Points take the new velocity (FLIP or APIC) and move with the grid, so material can flow, fold and separate.
The velocity gradient drives a Jaumann-rate stress update through the constitutive model, and the grid is cleared for the next step.
What it does
Models are staged in phases, as in a finite element analysis, and every phase runs on the same state on either engine.
2D plane strain or 3D sections extruded over depth ranges, terrain height grids from ESRI .asc files, and drape layers laid on the terrain.
MUSL or USL, uGIMP, linear or B-spline bases, FLIP or APIC transfer, cell-averaged volumetric strain against locking, CFL time step and Cundall damping.
Mohr-Coulomb with softening, Drucker-Prager, Hardening Soil, Soft Soil, fluid and linear elastic, drained, undrained or two-phase.
K0 or gravity start, region activation and excavation, rigid footings and walls, loads, and a strength-reduction search for the factor of safety.
Rayon on all cores in f64, or CUDA kernels compiled at run time with NVRTC in f32, chosen automatically from 2,000 points when a GPU is present.
Frames of displacement, velocity, stress, plastic strain and pore pressure, monitor and rigid-body series, PNG pictures and a binary results file.
Constitutive models
Mohr-Coulomb carries the flows. Hardening Soil and Soft Soil, new in this version, carry the stiffness before failure that settlement and excavation depend on, each with its cap calibrated on a simulated oedometer test, as PLAXIS does.
Non-associated, returned in principal stresses with plane, edge and apex returns, linear softening of c, φ and ψ to residual values, and a tension cut-off.
Matched to Mohr-Coulomb in plane strain, non-associated; with φ = 0 it is von Mises, which is Tresca with su = c in plane strain.
Schanz, Vermeer and Bonnier (1999): stress-dependent E50, Eoed and Eur with power m, hyperbolic shear hardening, Rowe dilatancy, a cap, and OCR or POP.
The PLAXIS model for soft clay: K = p'/κ*, a Cam-Clay cap hardening with λ* − κ*, Mohr-Coulomb failure, and M calibrated to K0nc. Takes λ*, κ* or Cc, Cs, e0.
A weakly compressible Newtonian fluid with p = K(1/J − 1), for dam breaks, flooding and impact on structures.
Drained, undrained in effective stress, or a single-point two-phase mixture with Darcy flow, for consolidation and saturated failure.
The element and solver-level tests above ran on the CPU. The Hardening Soil and Soft Soil GPU kernels reproduce the CPU element tests when compiled as host code in f32, and still have to be run on an NVIDIA device.
The application
A desktop application with an Input view for materials, regions, rigid bodies, loads and phases, and an Output view that plays the frames coloured by any of 13 quantities and charts any pair of series columns.
AI agent and MCP
The same 28 tools serve the embedded TigrimOS agent in the chat panel and any Model Context Protocol client, such as Claude Code, in m, kPa, kN/m³, degrees and seconds.
list_templatesnew_projectopen_projectsave_projectget_projectset_modeladd_materialupdate_materialadd_regionadd_rigidadd_loadadd_monitoradd_phase…validatecalculateget_resultslist_columnsget_curveshow_pictureget_particles
One demonstration, not an evaluation. Part of the 23 % excess came from the template's former damping of 0.05, which the agent reported but could not diagnose.
GPU engine
The CUDA kernels are compiled at run time with NVRTC, so one binary runs with or without a GPU. On the verification cases they give the CPU engine's results to the reported digits.
Verification and calibration
About 120 runs on one RTX 4090, documented in the paper with the CSV results of every run. Where the answers are off, the paper says why.
| Benchmark | Reference | TellusPM |
|---|---|---|
| Vibrating bar | Closed form | Second-order convergence with uGIMP |
| Finite-strain oedometer | Logarithmic-strain solution | Within 0.6 % |
| Terzaghi consolidation | Closed form | Within 3.9 % of the initial pore pressure |
| Sliding block | g(sin θ − μ cos θ) | Within 0.4 to 2.7 % |
| Griffiths and Lane slope | FE 1.35–1.40, Bishop 1.38 | Factor of safety 1.398 / 1.367 / 1.344 (cells 1 / 0.5 / 0.25 m) |
| Strip footing, Tresca clay | Prandtl 5.14 su | About 4 % above once local damping is off |
| Dam break | Martin and Moyce (1952) | Surge front within 1.4 to 3.1 % |
| Granular column collapse | Experiment, Lube's runout law | Squat 2D columns match; aluminium rods 25 % short; tall columns 1.3 to 1.7 times long |
| Patong debris flow, one source | Coupled DEM–SPH | 61 % of the moved mass within 150 m of Nanai Road, against 53 to 60 % |
| Patong, all FS < 1 cells | FE + DEM–SPH chain | 99th-percentile travel within 6 % (349 against 330 m); median 38 % longer |
| Hardening Soil, Soft Soil | Element tests, 1D settlement | E50, Eoed, K0, λ*, κ* and qf within 4 %; settlement within 2 % |
Case study
The 26 October 2026 Patong (Nanai) debris flow, run on GLO-30 terrain. Every 20 m cell whose saturated infinite-slope factor of safety is below 1 is released at once, with the strength reduced by γsat/γ' = 1.96.
A stress test of the code and a map of where debris collects, not a prediction of the event: the same rule also releases ground that did not fail.
Get started
A recent stable Rust toolchain builds everything. The CUDA engine needs only an NVIDIA driver and the NVRTC library of a CUDA 12 toolkit at run time.
# build and start cargo build --release -p telluspm ./target/release/telluspm # open a model and its results ./target/release/telluspm model.telluspm
# headless server on stdin/stdout cargo build --release -p telluspm-core ./target/release/telluspm-mcp # register with Claude Code claude mcp add telluspm -- \ /path/to/target/release/telluspm-mcp
# closed-form and element tests cargo test -p telluspm-mpm --release # CPU against CUDA, incl. HS and SS cargo run --release --example gpu_check all
column_collapsefootingslopeconsolidationslope_runoutfooting_3dcolumn_3dKnown limitations
Contact
Write to the TellusPM team about the method, a model that misbehaves, collaboration or licensing. A bug report is most useful with the project file and the version you ran.