Pressnet_plus

Staged multi-model learned surrogate for press forming structural simulation.

Pressnet++ is the code release accompanying the paper:

PRESSNET: A Forming Dataset for Structural Simulation in Pressed Blanks with Deep Learning Benchmarks — Panta et al., ASME IDEC/CIE 2025, Anaheim, CA (IDETC2025-163821).

What it does

Numerical simulation is integral to engineering design, replacing costly physical prototyping with virtual iterations; yet high-fidelity methods such as the finite element method (FEM) are computationally expensive, limiting how often they can run inside a design loop. Press-forming is an especially demanding case, because a single solve couples large visco-elastic deformation, moving-boundary contact, stress relaxation, and thermal diffusion in one transient, multi-body analysis.

Learned surrogates promise a faster alternative, but existing forming surrogates target a single field, a quasi-static setting, or a thermal-only response; none reproduces the complete multi-stage, multi-physics forming trajectory. PressNet++ closes this gap using the PressNet dataset of 150 transient pressed-forming trajectories (see the paper citation below for the canonical reference).

Contributions

• We propose PressNet++, a staged multi-model surrogate architecture that partitions the forming trajectory into specialized networks for the pressing, dwell, and release regimes, together with a parallel network for thermal diffusion, and show that it outperforms a monolithic single-model baseline.
• We provide a controlled benchmark of five architectures spanning two families — graph neural networks and a transformer-based PDE solver — on the same transient, multi-physics, large-deformation task.
• We evaluate generalization and stability through in-distribution, parametric-extrapolation, and unseen-shape splits, and through a one-step versus rolled-out comparison that isolates error accumulation. Global-attention models are more accurate on geometries seen during training, whereas local mesh-based message passing generalizes better to unseen geometries.
• We report architecture-specific hyperparameter ablations and a mesh-resolution study that includes cross-resolution transfer between coarse and fine meshes.

Architecture

PressNet++ is a staged multi-model architecture in which separate networks handle the pressing, dwell, and release regimes, and a parallel network handles thermal diffusion. We validated this design across five architectures spanning two families — graph neural networks (GCN, Reg-DGCNN, MeshGraphNet, Dilated-DGCNN) and the transformer-based solver Transolver.

Key findings:

• The staged formulation outperforms a monolithic baseline for every architecture. For Transolver, rolled-out y-displacement nRMSE drops from 6.57% to 3.20% in the coarse mesh and to 1.76% in the fine mesh.
• Architectures behave very differently out of distribution: Transolver is most accurate on die shapes seen in training, yet on unseen shapes MeshGraphNet is far more robust (9.47% versus 54.76% nRMSE).
• One-step accuracy below 0.4% displacement nRMSE for every model shows long-horizon stability issues and potential improvement through stabilization of autoregressive inference.
• Stress prediction has a much higher error than displacement in all scenarios.

Results

Training loss

Inference rollout

Quick start

conda env create -f environment.yml
conda activate pressnetpp

Training and inference run on GPU by default (torch is CUDA-built and the furthest_point_sampling extension used by dilated_dgcnn is CUDA-only). All device selection is automatic via torch.device("cuda" if torch.cuda.is_available() else "cpu").

Each architecture is trained three times — once per stage (Press / Dwell / Release) — then stitched together at inference. Replace --model with the architecture you want.

Training (one command per stage per model)

# --- Graph neural networks ---
python -m pressnetpp.train --config configs/train_dilated_dgcnn.json --stage 1 --model dilated_dgcnn
python -m pressnetpp.train --config configs/train_transolver.json  --stage 1 --model transolver
python -m pressnetpp.train --config configs/train_gcn.json          --stage 1 --model gcn

# Stage 2 (Dwell) and Stage 3 (Release) use the same commands with --stage 2 / --stage 3

Inference (3-stage rollout)

python -m pressnetpp.inference --config configs/inference_multi.json

Set checkpoint_path_1/2/3 in the config to the three trained stage checkpoints (Press / Dwell / Release). core_model can be a single string (same architecture all stages), a list of three, or a dict with stage_1 / stage_2 / stage_3 keys.

Verified

The package was smoke-tested end-to-end. All device logic uses torch.device("cuda" if torch.cuda.is_available() else "cpu"), so the code runs on GPU when one is present (torch is CUDA-built and the furthest_point_sampling extension is CUDA-capable).

• import pressnetpp.train, pressnetpp.inference ✓
• TrajectoryDataset loads the real coarse dataset (13,694 samples, 887 nodes, all expected fields) ✓
• One real training step (forward + backward + optimizer step) on real data passes for encode_process_decode, gcn, transolver, regDGCNN_seg, and regpointnet_seg ✓
• dilated_dgcnn requires a GPU — its furthest_point_sampling extension has no CPU fallback; it runs on GPU through the same code path.

Repository layout

Pressnet++/
  README.md
  LICENSE
  CITATION.bib
  environment.yml
  configs/                       # example JSON configs
  pressnetpp/
    train.py                     # train one stage (Press / Dwell / Release)
    inference.py                 # run the full 3-stage rollout
    models/                      # 5 benchmarked architectures + wrapper
    utilities/                   # dataset loader, eval, plot, anim, paraview

See configs/*.json for full hyperparameter sets; paths are user-configurable.

Data

The PressNet dataset (150 trajectories × 15 die shapes × 10 geometric variations, coarse/medium/fine meshes, 1500 time steps) is not redistributed here. The dataset and its download instructions are part of the parent PressNet repository (see the ASME paper citation below for the canonical reference). The thermal sub-dataset (datasets/data/thermal/s_quarter_1500_withthermal.h5) lives in the parent repo.

Note on thermal: the parallel thermal-diffusion network referenced in the paper is not included in this release. The thermal dataset is present, but the thermal model and its evaluation utilities live in a separate codebase.

Citation

If you use this code or dataset, please cite:

@inproceedings{PRESSNET2025,
  author    = {Prince Panta and Saroj Belbase and Rachit Rijal and Bipin Shrestha and Nirmal Prasad Panta and Dikshya Parajuli and Rujal Acharya and Saugat Kafley and Amit Regmi and Ken Igeta and Akio Tanaka and Christopher McComb},
  title     = {{PRESSNET: A Forming Dataset for Structural Simulation in Pressed Blanks with Deep Learning Benchmarks}},
  booktitle = {Proceedings of the ASME 2025 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (IDETC/CIE 2025)},
  year      = {2025},
  address   = {Anaheim, CA, USA},
  paperid   = {IDETC2025-163821},
  publisher = {American Society of Mechanical Engineers (ASME)},
  doi       = {10.1115/DETC2025-163821},
  url       = {https://doi.org/10.1115/DETC2025-163821}
}

License

Copyright Accelerated Computing Pvt. Ltd. See LICENSE.