jaewan-wod33/Single_vs_Multiple_Branches_in_S_DeepONet

Datasets — Single vs. Multiple Branches in DeepONet and S-DeepONet

Finite-element datasets accompanying the paper "Single vs. Multiple Branches in DeepONet and S-DeepONet: Network Architecture Follows Coupling in Multiphysics Systems" (Park, Kobayashi, Liu, Koric, Abueidda, Alam — arXiv:2507.03660).

• 📄 Paper: https://arxiv.org/abs/2507.03660
• 💻 Code: https://github.com/benjamin0303/Single_vs_Multiple_Branches_in_S_DeepONet

These are the full-field FE solution datasets used to compare single-branch (1br) and multi-branch / MIONet (2br) (S-)DeepONet architectures across multiphysics problems with different coupling. (The third problem in the paper, single-physics reaction–diffusion, is not hosted here — it is regenerated from the data/generation/ scripts in the code repo.)

Structure

thermo_electrical/
├── coupled/      [phi(t)]_coupled_Qrhoe(t).npz
└── uncoupled/    elec[phi(t)]_thermal_t_uncouple_thermal.npz
                  elec[phi(t)]_thermal_t_uncouple_elet.npz
thermo_mechanical/
├── coupled/      xy_train_testing.npy  flux_filtered.npy  disp_filtered.npy
│                 filtered_temp_data.npy  filtered_stress_data.npy
└── uncoupled/    xy_train_testing.npy  flux_filtered.npy  disp_N_train.npy
                  filtered_temp_data.npy  stress_time_steps_3000.npy

Each multiphysics problem has a coupled and an uncoupled variant (data from fully-coupled vs. independently-solved FE analyses). A small demo/ subset (first samples of each case) is provided under every folder for quick testing without downloading the full arrays.

Array keys / shapes

• Thermo-electrical, coupled — a single file [phi(t)]_coupled_Qrhoe(t).npz (the two fields are solved together, so both inputs and both outputs are bundled in one array each): grid_input (nodes, 2); input_Qext_rhoe (N, steps, 2) with [...,0]=Q (heat source), [...,1]=ρₑ (resistivity); target_T_phi (N, nodes, 2) with [...,0]=T (temperature), [...,1]=φ (potential).
• Thermo-electrical, uncoupled — two files (thermal & electrical solved independently): ..._thermal.npz: x_grid, t_grid, Q_ext_all, T_solutions; ..._elet.npz: rho_e_all, phi_solutions.
• Thermo-mechanical (.npy per array): nodal coords xy_train_testing, input amplitudes (flux_*, disp_*), and last-frame temperature / stress fields along the solidification slice.

Usage

Download the whole dataset (or a single case) and load with NumPy:

from huggingface_hub import snapshot_download, hf_hub_download
import numpy as np

REPO = "jaewan-wod33/Single_vs_Multiple_Branches_in_S_DeepONet"

# one case (small) — thermo-mechanical coupled coordinates
xy = np.load(hf_hub_download(REPO, "thermo_mechanical/coupled/xy_train_testing.npy", repo_type="dataset"))

# a quick-start demo subset (no need to pull the multi-GB arrays)
snapshot_download(REPO, repo_type="dataset", allow_patterns="thermo_mechanical/coupled/demo/*",
                  local_dir="data_demo")

# the full thermo-electrical coupled file
f = hf_hub_download(REPO, "thermo_electrical/coupled/[phi(t)]_coupled_Qrhoe(t).npz", repo_type="dataset")
d = np.load(f)
print(d.files)

To run the training scripts in the code repo, place these files under data/<problem>/<coupling>/ (the script paths are relative to the repo root — see the repo's data/README.md).

Note on size. The temperature/stress arrays are large (~14–27 GB each; ~98 GB total). Use the demo/ subsets or hf_hub_download per file to avoid pulling everything.

License & citation

Released under the MIT License.

@article{park2026branches,
  title   = {Single vs. Multiple Branches in DeepONet and S-DeepONet: Network Architecture Follows Coupling in Multiphysics Systems},
  author  = {Park, Jaewan and Kobayashi, Kazuma and Liu, Qibang and Koric, Seid and Abueidda, Diab and Alam, Syed Bahauddin},
  journal = {arXiv preprint arXiv:2507.03660},
  year    = {2026},
  url     = {https://arxiv.org/abs/2507.03660}
}

Data generated with finite-element simulations on the NCSA Delta / DeltaAI systems, University of Illinois Urbana-Champaign.