GeoLGC-Dataset: Processed Microchannel Cold-Plate Topologies and Flow-Response Labels

# GeoLGC-Dataset Processed microchannel cold-plate topologies and their flow-response labels used in the study "GeoLGC-Net: Geometry-Aware Local-Global Consistent Multi-Task Surrogate Model for Non-Parametric Microchannel Topology Evaluation". ## Files - `dataset.h5`: HDF5 file containing all released samples (fields described below). - `LICENSE.txt`: license terms for the dataset. - `README.md`: this document. ## Dataset content `dataset.h5` contains one group per split: `train`, `val`, and `test`. Each group stores the following datasets: | Field | Shape | Dtype | Description | | --- | --- | --- | --- | | `sample_id` | (N,) | int | Unique sample identifier. | | `binary_topology` | (N, 128, 128) | uint8 | Binary topology; 0 = fluid, 1 = solid. | | `velocity_magnitude` | (N, 128, 128) | float32 | Steady-state velocity magnitude field, defined on all nodes (zero in solid nodes). | | `low_velocity_mask` | (N, 128, 128) | uint8 | Low-velocity (dead-zone) mask; 1 = fluid node with velocity below threshold, 0 otherwise (including all solid nodes). | | `pressure_drop` | (N,) | float64 | Pressure drop across the channel, computed as `c_s^2 * (mean(rho_in) - rho_out)` from the steady state. | The arrays are stored in row-major order, with array index `(row, col)` corresponding to the physical `(y, x)` node position. ## Conventions - **Binary topology**: 0 = fluid, 1 = solid. The topology is the raw geometric input; the signed distance field (SDF) is **not** included, because it can be computed from the binary topology with the formula given in the paper: `S(x) = d(x, Omega_s) - d(x, Omega_f)`. Normalized x/y coordinate channels are also derivable directly from array indices. - **Velocity magnitude**: the paper model predicts the velocity *magnitude* field; the raw velocity components are therefore not released. - **Low-velocity mask**: a fluid node is marked as low-velocity when `v(x) < 0.05 * v_max`, where `v_max` is the maximum velocity magnitude over fluid nodes of the same sample. - **Quality control**: every released sample passed the steady-state convergence check and the quality checks (connectivity, throat width, flow cross-section); divergent or non-converged samples are excluded. - **No code included**: this package does not contain the topology generator, the GPU-LBM solver, or the model training implementation. ## Notes on scope The public release is intended to support subsequent topology-flow surrogate studies rather than exact reproduction of the complete computational pipeline. The complete candidate pool, fixed experimental manifests, the topology generator, the GPU-LBM solver, and the model training implementation are not included. ## Regeneration script The script `Scripts/build_public_dataset.py` in the project rebuilds `dataset.h5` from the local processed data. It requires `h5py`. ```bash pip install h5py python Scripts/build_public_dataset.py ```