SELTO Dataset

A Benchmark Dataset for Deep Learning-based Methods for 3D Topology Optimization. One can find a description of the provided dataset partitions in Section 3 of Dittmer, S., Erzmann, D., Harms, H., Maass, P., SELTO: Sample-Efficient Learned Topology Optimization (2022) https://arxiv.org/abs/2209.05098. <br> Every dataset container consists of multiple enumerated pairs of CSV files. Each pair describes a unique topology optimization problem and a corresponding binarized SIMP solution. Every file of the form {i}.csv contains all voxel-wise information about the sample i. Every file of the form {i}_info.csv file contains scalar parameters of the topology optimization problem, such as material parameters. <br> This dataset represents topology optimization problems and solutions on the bases of voxels. We define all spatially varying quantities via the voxels' centers -- rather than via the vertices or surfaces of the voxels.<br> In {i}.csv files, each row corresponds to one voxel in the design space. The columns correspond to ['x', 'y', 'z', 'design_space', 'dirichlet_x', 'dirichlet_y', 'dirichlet_z', 'force_x', 'force_y', 'force_z', 'density']. x, y, z - These are three integer indices stating the index/location of the voxel within the voxel mesh. design_space - This is one ternary variable indicating the type of material density constraint on the voxel within the TO problem formulation. "0" and "1" indicate a material density fixed at 0 or 1, respectively. "-1" indicates the absence of constraints. dirichlet_x, dirichlet_y, dirichlet_z - These are three binary variables defining whether the voxel contains homogenous Dirichlet constraints in the respective axis direction. force_x, force_y, force_z - These are three floating point variables giving the three spacial components of the forces applied to each voxel. All forces are body forces given in [N/m^3]. density - This is a binary variable stating whether the voxel carries material in the solution of the topology optimization problem. Any of these files with the index i can be imported using pandas by executing: <pre><code class="language-python">import pandas as pd directory = ... file_path = f'{directory}/{i}.csv' column_names = ['x', 'y', 'z', 'design_space','dirichlet_x', 'dirichlet_y', 'dirichlet_z', 'force_x', 'force_y', 'force_z', 'density'] data = pd.read_csv(file_path, names=column_names)</code></pre> From this pandas dataframe one can extract the torch tensors of forces F, Dirichlet conditions ω_Dirichlet, and design space information ω_design using the following functions: <pre><code class="language-python">import torch def get_shape_and_voxels(data): shape = data[['x', 'y', 'z']].iloc[-1].values.astype(int) + 1 vox_x = data['x'].values vox_y = data['y'].values vox_z = data['z'].values voxels = [vox_x, vox_y, vox_z] return shape, voxels def get_forces_boundary_conditions_and_design_space(data, shape, voxels): F = torch.zeros(3, *shape, dtype=torch.float32) F[0, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['force_x'].values, dtype=torch.float32) F[1, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['force_y'].values, dtype=torch.float32) F[2, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['force_z'].values, dtype=torch.float32) ω_Dirichlet = torch.zeros(3, *shape, dtype=torch.float32) ω_Dirichlet[0, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['dirichlet_x'].values, dtype=torch.float32) ω_Dirichlet[1, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['dirichlet_y'].values, dtype=torch.float32) ω_Dirichlet[2, voxels[0], voxels[1], voxels[2]] = torch.tensor(data['dirichlet_z'].values, dtype=torch.float32) ω_design = torch.zeros(1, *shape, dtype=int) ω_design[:, voxels[0], voxels[1], voxels[2]] = torch.from_numpy(data['design_space'].values.astype(int)) return F, ω_Dirichlet, ω_design</code></pre> The corresponding {i}_info.csv files only have one row with column labels ['E', 'ν', 'σ_ys', 'vox_size', 'p_x', 'p_y', 'p_z']. E - Young's modulus [Pa] ν - Poisson's ratio [-] σ_ys - Yield stress [Pa] vox_size - Length of the edge of a (cube-shaped) voxel [m] p_x, p_y, p_z - Location of the root of the design space [m] Analogously to above, one can import any {i}_info.csv file by executing: <pre><code class="language-python">file_path = f'{directory}/{i}_info.csv' data_info_column_names = ['E', 'ν', 'σ_ys', 'vox_size', 'p_x', 'p_y', 'p_z'] data_info = pd.read_csv(file_path, names=data_info_column_names)</code></pre>