⚡ electrical engineering

GPU Acclerated Explicit FEM–Level-Set DEM Coupling for Rigid–Soft Mixtures with LLM-Assisted Model Construction

This paper introduces an open-source, GPU-accelerated framework that couples explicit FEM and level-set DEM to simulate soft-rigid particle mixtures with a conservative contact model, while also providing an LLM-assisted platform to streamline model construction and verification.

Original authors: Yihao Shi

Published 2026-09-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yihao Shi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where the ground beneath your feet is not just a uniform surface, but a chaotic mixture of hard, jagged stones and soft, squishy rubber balls. This is the reality of many natural and industrial materials, from sand mixed with clay to recycled tires blended with gravel. When these two very different types of particles interact, they do not behave like a simple average of their parts. The hard stones lock together to form a rigid skeleton, while the soft grains squeeze and stretch to fill the gaps, creating a complex, shifting network of forces. To understand how such mixtures flow, settle, or support weight, scientists must track two things simultaneously: the precise motion of every hard particle and the internal stretching and stress within every soft particle. For decades, simulating this dual nature on a computer has been a formidable challenge, often requiring so much computing power that realistic models were impossible to run.

A new approach developed by independent researcher Yihao Shi offers a solution by merging two powerful simulation techniques into a single, high-speed system. The method treats hard, irregular particles using a technique that maps their exact 3D shapes, while soft, deformable particles are broken down into tiny tetrahedral blocks that can stretch and compress like real rubber. The innovation lies in how these two distinct worlds talk to each other. The system uses a specialized, energy-conserving rule to calculate the forces when a soft block bumps into a hard shape, ensuring that the work done by the collision matches the energy stored inside the material. This prevents the simulation from creating or losing energy artificially, a common problem in computer models that can lead to unrealistic results. By running these calculations on a graphics processing unit, the same hardware found in modern video games, the researcher achieved a speedup that makes it possible to simulate thousands of these mixed particles interacting with high efficiency.

To make this complex technology accessible, the researcher also built an assistant powered by a large language model, a type of artificial intelligence trained on vast amounts of text and code. Instead of requiring a user to write hundreds of lines of code to set up a simulation, a person can now describe a scenario in plain English or upload a 3D model from a design program, and the system automatically generates the necessary geometry, checks the quality of the internal mesh, and sets the physical parameters. If the simulation fails or produces unstable results, the system can automatically diagnose the issue, adjust the settings, and restart the calculation without human intervention. This creates a seamless workflow from a simple idea to a verified physical result, lowering the barrier for scientists and engineers to study these complex materials.

The power of this new framework was tested through a series of rigorous experiments. First, the system was asked to replicate a classic physics problem: a soft sphere pressing against a flat surface. The simulation matched the theoretical predictions for how the sphere deforms and how the pressure spreads across the contact area, proving that the underlying math was correct. Next, the researcher simulated collisions between soft and hard objects, including spheres and cubes, to ensure that momentum and energy were conserved perfectly during impact. The results showed that the system could handle the transfer of energy from motion into deformation and back again with extreme precision, even when the shapes were irregular and the collisions were off-center.

The framework was then pushed to its limits by simulating a large-scale event: a funnel discharging a mixture of four hundred soft grains and four hundred hard, angular particles. As the gate opened, the mixture flowed out, with the soft grains deforming to fit through the narrow opening while the hard grains tumbled over them. The simulation tracked the energy of the entire system, showing how gravitational potential energy converted into motion, then into the elastic stretching of the soft particles, and finally dissipated as heat through friction and internal damping. The model successfully captured the collective behavior of the mixture, demonstrating that the soft and hard phases influenced each other in ways that simpler models could not predict.

Finally, the researcher tested the system's ability to handle extreme compression. In one scenario, a box filled with a mix of soft and hard spheres was squeezed from all sides. As the walls moved inward, the soft particles flattened and distorted significantly, while the hard particles rearranged to fill the remaining space. The simulation revealed that when the mixture contained a high proportion of soft particles, the intense local deformation at the contact points between soft and hard grains caused the internal mesh of the soft particles to distort severely, eventually limiting how much the material could be compressed. This finding highlights a physical limit in the behavior of such mixtures that would be invisible to a model that treats all particles as rigid.

Throughout these tests, the system demonstrated remarkable efficiency. In the funnel discharge experiment, the contact force calculations—the most computationally expensive part of the process—accounted for more than three-quarters of the total computing time, yet the system still managed to process hundreds of steps per second on a single graphics card, with throughput decreasing from approximately 202 steps per second for small systems to about 89 steps per second for systems with 1,200 particles. The memory usage scaled predictably, growing from less than one gigabyte for a small number of particles to just over twelve gigabytes for a system with twelve hundred particles. This performance confirms that the method is not only physically accurate but also practical for studying large, complex systems.

The integration of the large language model assistant proved to be a critical component of the workflow. By connecting natural language requests and 3D design files to the simulation engine, the system eliminated the tedious and error-prone steps of manual setup. The assistant could generate the necessary geometry, create the internal mesh, and verify that the physical parameters were within a stable range before the simulation even began. If a problem arose, such as an unstable time step or a mesh quality issue, the system could identify the specific setting causing the trouble and correct it automatically. This level of automation suggests a future where complex physical simulations are as easy to run as a standard software application, allowing researchers to focus on the science rather than the mechanics of the code.

The study concludes that this coupled approach successfully bridges the gap between the rigid world of discrete particles and the continuous world of deformable materials. It provides a tool that can resolve the intricate details of how soft grains deform while interacting with hard obstacles, all while maintaining the conservation of energy and momentum. While the current method faces challenges with extreme local distortion that can break the internal mesh of the soft particles, the results demonstrate a significant step forward in our ability to model the complex, mixed materials that surround us. The open-source nature of the software ensures that other researchers can build upon this foundation, potentially leading to new insights into soil mechanics, pharmaceutical manufacturing, and the processing of recycled materials.

Technical Summary: GPU-Accelerated Explicit FEM–Level-Set DEM Coupling for Rigid–Soft Mixtures with LLM-Assisted Model Construction

Problem Statement

Rigid–soft granular mixtures, found in applications such as sand–rubber composites and compacted assemblies, exhibit complex mechanical behaviors that cannot be captured by simple interpolations between purely rigid or purely soft material models. Their response depends on the connectivity of rigid contact networks, the deformation of soft grains, and the evolution of mixed contacts. Existing numerical methods face a trade-off:

Discrete Element Method (DEM): Efficient for rigid grains but relies on simplified geometries (spheres, clusters) or expensive surface meshes. It cannot resolve internal deformation or stress within soft particles.
Finite Element Method (FEM): Capable of resolving internal deformation and stress in soft particles but requires volume discretization and complex contact handling, making it computationally expensive for large systems with many particles.
Level-Set DEM (LSDEM): Handles irregular rigid geometries efficiently via signed-distance fields (SDF) but lacks the ability to model soft particle deformation.

The core challenge is to develop a coupled framework that simultaneously resolves the motion of irregular rigid particles and the large-strain deformation of soft particles within a single, computationally efficient simulation, while ensuring physical consistency in energy and momentum transfer. Additionally, setting up such complex coupled simulations typically requires significant manual effort in geometry, meshing, and parameter selection.

Methodology

The author presents a GPU-accelerated explicit coupling of the Finite Element Method (FEM) and Level-Set Discrete Element Method (LSDEM), integrated into the open-source GeoTaichi platform.

1. Physical Formulation

Rigid Phase (LSDEM): Rigid particles are represented by signed-distance fields (SDF) in body-local coordinates. This allows for the treatment of arbitrary shapes (angular, concave, image-derived) without separate contact procedures for each geometry. Rigid body motion is governed by linear and angular momentum balance equations.
Soft Phase (FEM): Soft particles are modeled as deformable continua using a compressible Neo-Hookean material law. They are discretized using first-order tetrahedral meshes. The motion is governed by the total-Lagrangian momentum balance, solved via explicit time integration.
Coupling Strategy: The two phases interact through an explicit contact law. The solver calculates contact candidates, evaluates interface-specific gaps, applies contact forces, and transfers equal and opposite forces to both phases in a single step without iterative interface solving.

2. Contact Detection and Force Model

Search Algorithm: A two-stage contact search is implemented on the GPU using a Bounding-Volume Hierarchy (BVH).
- Broad Phase: Identifies nearby bodies or surface primitives.
- Narrow Phase: Evaluates specific gaps for three interaction types:
  1. FEM–FEM: Point–triangle and edge–edge tests on triangulated surfaces.
  2. FEM–LSDEM: Projection of FEM surface nodes onto the rigid body's SDF.
  3. LSDEM–LSDEM: Sampling rigid body surfaces against the target's SDF.
Energy-Consistent Contact Model: A conservative elastic contact model is derived from a common potential to ensure consistency between contact work and stored elastic energy.
- Normal forces follow a power-law potential (quadratic for Hertz-like behavior).
- Tangential forces use an elastic trial branch with Coulomb friction limits.
- Viscous damping and frictional dissipation are accounted for separately to maintain energy balance.
- Surface factors ( $A_c$ ) are introduced to normalize contact forces based on the discretization of the interacting features (vertices, edges, or surface samples).

3. LLM-Assisted Workflow

To address the complexity of model setup, the author introduces a Large Language Model (LLM)-assisted workflow connected via the Model Context Protocol (MCP).

Input: Users can provide natural language descriptions or Blender scenes.
Process: The LLM retrieves relevant physical representations, constructs geometries (using tools like Hunyuan3D-2 and fTetWild), generates tetrahedral meshes, and configures material/solver parameters.
Verification: Deterministic tools, rather than the LLM, evaluate mesh quality, timestep stability, and physical consistency (e.g., energy balance).
Correction: If criteria are not met, the system identifies the specific setting to revise (e.g., reducing timestep, fixing initial overlaps) and re-runs the simulation automatically.

Key Contributions

GPU-Accelerated FEM–LSDEM Coupling: A novel method for simulating soft–rigid mixtures that combines tetrahedral FEM for deformable grains and SDF-based LSDEM for irregular rigid grains. The framework is implemented in the open-source GeoTaichi platform.
Energy-Conserving Contact Formulation: A contact model derived from a common potential that ensures consistency between contact work and stored energy, handling FEM–FEM, FEM–LSDEM, and LSDEM–LSDEM interactions with appropriate surface weighting.
LLM-Assisted Modeling Platform: An integrated workflow using MCP to connect natural language/Blender inputs to model generation, execution, physical verification, and post-processing, reducing the barrier to entry for complex multiphysics simulations.

Results and Verification

The framework was validated through six numerical examples:

Geometry Generation: Successfully converted text descriptions and single images (e.g., sand grains, dragons) into watertight FEM volumes and tetrahedral meshes for gravity-driven simulations.
Hertz Contact: A soft sphere indenting a rigid plane showed convergence toward analytical Hertz solutions. Refining the mesh reduced the relative $L_2$ error in pressure profiles from 7.92% to 5.27% and loading response error from 14.53% to 3.78%.
Collision Tests: Frictionless collisions between sphere–sphere, sphere–cube, and cube–cube pairs demonstrated excellent conservation of linear and angular momentum (residuals $\approx 10^{-15}$ to $10^{-12}$ ) and energy balance, with temporal errors decreasing as the timestep was refined.
Inclined-Plane Motion: Simulations of soft spheres and cubes sliding/rolling on an incline matched analytical solutions for various friction coefficients.
Funnel Discharge (Collective Response): A mixture of 400 soft and 400 rigid particles discharging through a funnel showed realistic collective behavior. Energy analysis confirmed that gravitational potential energy was correctly converted to kinetic, strain, and contact energy, with total balance residuals under 0.49%.
Isotropic Compaction: Simulations of 50% and 100% soft-particle packings under isotropic compression illustrated the framework's ability to resolve multi-body interactions and soft-grain deformation. The 50% mixture showed severe local mesh distortion at soft–rigid interfaces, limiting the compaction range.

Performance:

GPU profiling on an NVIDIA RTX 3080 Ti indicated that contact-force calculation dominates the computational cost (75.2% of sampled time), followed by neighbor search (6.7%) and FEM internal force evaluation (5.9%).
Throughput decreased from ~202 steps/s to ~89 steps/s as particle count increased from 5 to 1200, with GPU memory usage scaling from 0.56 GB to 12.24 GB.

Significance and Claims

The paper claims that this framework provides a robust, open-source solution for simulating soft–rigid mixtures where both particle shape and internal deformation are critical. By integrating BVH-accelerated contact detection and an energy-consistent contact model, the method maintains physical fidelity while leveraging GPU parallelism.

The author emphasizes that the LLM-assisted workflow significantly lowers the barrier to entry for setting up these complex simulations, automating the transition from conceptual input (text or 3D scenes) to verified physical models. However, the paper remains modest regarding current limitations:

The computational cost of contact-force evaluation remains the primary bottleneck.
Severe local deformation can lead to mesh distortion, limiting the accessible compaction range in highly mixed systems.
Future work is identified as focusing on reducing contact evaluation costs and improving the treatment of mesh distortion to extend the applicability of the method to more extreme deformation scenarios.