Accepted at NeurIPS 2026
- What is GMP?
GMP (Geometry Meets Physics) is a disk-data-free pre-training framework for neural surrogate models on unstructured 3D geometries, tailored to both steady-state and transient physical regimes.
- How is GMP different from existing PDE foundation models?
GMP addresses the core bottlenecks of data generation, storage, and transferability in unstructured physical modeling:
-
Pure Online Strategy (Disk-Data-Free): Rather than relying on massive, expensive pre-computed simulation data stored on disk (which often consume terabytes of storage and suffer from severe I/O bottlenecks), GMP generates pre-training data purely on-the-fly. This eliminates disk-space requirements and makes large-scale pre-training lightweight and accessible.
-
Dual-Regime Pre-Training: Tailored to the distinct nature of steady-state and time-dependent physical problems:
- Steady-state regime (Geometry-driven): GMP employs intrinsic shape descriptors (coordinates, curvatures, normals, and SGD) to learn rich representations of complex 3D domains without running expensive offline simulations.
- Transient regime (Physics-driven): GMP uses GPU-accelerated solvers to generate synthetic PDE dynamics online, enabling scalable pre-training for time-dependent evolutions without storage overhead.
-
Data-Efficient Downstream Transfer: Models pre-trained with GMP achieve faster convergence, greater data efficiency, and higher accuracy during fine-tuning, delivering higher performance gains under realistic, low-data regimes.
-
Architecture Versatility (Designed for GMP-Net, Proven on Baselines): The pre-training framework is specifically designed for GMP-Net (our proposed architecture derived from SMART [Hagnberger & Niepert, ICML 2026], featuring a multi-branch geometry-physics formulation with cross-modal latent attention). Beyond GMP-Net, we also demonstrate that GMP pre-training generalizes effectively to other popular unstructured architectures, including Transolver++ and AB-UPT.
-
Phased Fine-Tuning & Warmup Adaptation: Integrates structured fine-tuning schedules that freeze the pre-trained backbone while adapting task-specific projection heads and global context modulators, transitioning smoothly to full end-to-end training.
Luis Medrano-Navarro, Giacomo Baldan, Qiang Liu, Benjamin Holzschuh, Jan Hagnberger, Mathias Niepert, and Nils Thuerey
Abstract: Neural surrogate models for Partial Differential Equations (PDEs) on unstructured 3D geometries are often limited by poor generalization and the high cost of generating large-scale training datasets. Consequently, pre-training on massive datasets of related PDE dynamics has emerged as a critical alternative to enhance the robustness and scalability of these models. However, this strategy is neither compute- nor data-efficient, as it relies on massive pre-computed data that is very costly to generate. In this work, we introduce a disk-data-free pre-training framework tailored to both steady-state and transient regimes. For steady-state problems, we propose a geometry-driven strategy that leverages intrinsic shape descriptors to learn representations of complex 3D domains. For transient problems, we introduce a physics-driven approach based on online generation of synthetic PDE data, enabling scalable pre-training without reliance on expensive datasets. Across multiple experiments, our approach achieves faster convergence, greater data efficiency, and higher accuracy during fine-tuning, particularly under realistic low-data regimes. This methodology provides a practical pathway toward data-efficient neural emulators for large-scale simulations.
Read from: [arXiv:2610.03363]
Cite as:
@inproceedings{medrano2026geometry,
title={Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers},
author={Medrano-Navarro, Luis and Baldan, Giacomo and Liu, Qiang and Holzschuh, Benjamin and Hagnberger, Jan and Niepert, Mathias and Thuerey, Nils},
booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
year={2026},
url={https://openreview.net/forum?id=jidnQA2TIz}
}GMP is evaluated across six diverse downstream engineering benchmarks on unstructured surface and volumetric meshes, covering both steady-state and transient physical regimes:
| Benchmark | Physical Domain | Mesh Type | Target Output Fields | Description |
|---|---|---|---|---|
| DrivAerML | Automotive Aerodynamics | 3D Surface & Volume | Surface |
500 geometric variations of the DrivAer notchback car model under fixed inlet velocity ( |
| SHIFT-Wing | Aerospace Aerodynamics | 3D Surface & Volume | Surface |
Parametrically perturbed NASA Common Research Model (CRM) configurations under systematically varying operating conditions (Mach, Reynolds number, angle of attack). |
| SHIFT-Crash | Structural Mechanics | 3D Surface | Displacement |
Direct mapping from initial vehicle geometry and crash parameters to the final deformed state under steady-state formulation. |
| Benchmark | Physical Domain | Mesh Type | Rollout Horizon | Description |
|---|---|---|---|---|
| KuramotoβSivashinsky (KS) | Chaotic Nonlinear Dynamics | 2D Deformed Grid | 30 steps | 2D chaotic spatiotemporal dynamics interpolated onto irregular grids. Tests cross-dimensional generalization (3D |
| Flow Around Ellipse | Fluid Mechanics (Unsteady CFD) | 2D Unstructured Mesh | 100 steps | Incompressible 2D flow past an elliptical obstacle, predicting evolving velocity and pressure fields. |
| SHIFT-Crash | Dynamic Impact Mechanics | 3D Surface | 12 steps | Autoregressive temporal rollout across 12 discrete time steps capturing high-rate nonlinear deformation dynamics and stress distribution during vehicle collision. |
Clone the repository and set up the Python 3.12 environment:
# Create and activate environment
conda create -n gmp python=3.12 -y
conda activate gmp
# Install dependencies
pip install -r requirements.txtPre-trained model weights for GMP-Net and baselines are hosted on Hugging Face:
Download the pre-trained checkpoints and place them in the corresponding outputs_pretraining/ subfolders:
| Model / Regime | Pre-Training Stage | Destination Checkpoint Path |
|---|---|---|
| GMP-Net | Geometry Pre-Training (Steady) | outputs_pretraining/gmpnet/epoch-200.pth |
| GMP-Net | Physics Pre-Training (Transient) | outputs_pretraining/gmpnet_physics/epoch-199-EMA.ckpt |
| AB-UPT | Geometry Pre-Training | outputs_pretraining/abupt/epoch-200.pth |
| Transolver++ | Geometry Pre-Training (Surface) | outputs_pretraining/transolver_pp_surface/epoch-200.pth |
| Transolver++ | Geometry Pre-Training (Volume) | outputs_pretraining/transolver_pp_volume/epoch-200.pth |
Expected folder hierarchy:
outputs_pretraining/
βββ gmpnet/
β βββ epoch-200.pth
βββ gmpnet_physics/
β βββ epoch-199-EMA.ckpt
βββ abupt/
β βββ epoch-200.pth
βββ transolver_pp_surface/
β βββ epoch-200.pth
βββ transolver_pp_volume/
βββ epoch-200.pth
Once weights are placed in their respective folders, fine-tuning can be launched directly without any path changes.
Downstream fine-tuning and evaluation runs are configured with Hydra. Execute training commands from the root repository directory:
# Steady-state fine-tuning (e.g., DrivAerML)
python -m src.train --config-name=config_gmpnet_drivaerml_pt
# Transient autoregressive fine-tuning (e.g., KuramotoβSivashinsky)
python -m src.train --config-name=config_gmpnet_ks_pt
# Training from scratch
python -m src.train --config-name=config_gmpnet_drivaerml_scratchConfigurations for all other benchmarks and baseline models (config_abupt_*, config_transolverpp_*) are located in config/.
Weights & Biases (wandb) is integrated for tracking experiments:
- Default entity: Set to
null, meaning runs automatically log to your personal logged-in WandB account. - Custom entity / team: Specify your entity at runtime:
python -m src.train --config-name=config_gmpnet_drivaerml_pt wandb.entity=your_team_or_user
- Disable WandB or run offline:
python -m src.train --config-name=config_gmpnet_drivaerml_pt wandb.setup.mode=disabled
βββ config/ # Hydra configuration tree
β βββ callbacks/ # Callback configs (finetuning freeze schedule, metrics)
β βββ data/ # Dataset configs (drivaerml, crasht, ks)
β βββ model/ # Model architecture and checkpoint mapping configs
β βββ trainer/ # PyTorch Lightning trainer defaults
β βββ wandb/ # Weights & Biases experiment tracking configs
βββ outputs_pretraining/ # Pre-trained model checkpoints (download from Hugging Face)
βββ src/
β βββ callbacks/ # FineTuneCallback, ModelStats, CarMetrics
β βββ models/
β β βββ gmpnet/ # GMP-Net architecture and wrappers
β β βββ baselines/ # Transolver++, AB-UPT, and SMART baselines
β β βββ fft/ # Full fine-tuning (FFT) weight-loading wrappers
β βββ train.py # Downstream training entry point
βββ requirements.txt
This project is part of research in physics-based deep learning at the Physics-based Simulation group at the Technical University of Munich (TUM).
This project is licensed under the Apache 2.0 License.

