Dataset for "Physics-Aware Multi-Modal Learning for Real-Time Working Distance Prediction in Laser-Wire Directed Energy Deposition"

Overview This dataset contains the preprocessed model-input data used to train and evaluate PAM-Net, a physics-aware multi-modal network for real-time working distance (WD) prediction in laser-wire based Directed Energy Deposition (LW-DED). The data were acquired using the meltWatch process monitoring system and support the findings reported in the associated publication. The dataset provides train/validation/test splits and the fitted normalization scalers required to reproduce the model inputs and to inverse-transform model outputs to physical units. Folder structure model_input_data/├── train/ per-experiment .npz ┐├── val/ per-experiment .npz ├ train : val : test = 8 : 1 : 1├── test/ per-experiment .npz ┘└── scaler/ fitted normalization parameters Each .npz corresponds to one deposition experiment. File names denote the deposition pattern and, for line patterns, the build direction: Overlap_Nlayer, Spiral, Zigzag_x (x-direction build), Zigzag_y (y-direction build). Contents of each .npz Each sample corresponds to one distance-domain window. Arrays (N = number of windows in the file): - x_img : (N, 30, 112, 112) float32 — melt-pool image sequence, 30 frames, 112x112, grayscale, normalized to [0, 1].- x_sensor : (N, 50, 7) float32 — 7 measured sensor channels, 50-step window, stored as raw values (normalized at load time via the scaler). Channels: [1] Real laser power (R_LP) [2] Real robot speed (R_RS) [3] Real wire feed speed (R_WS) [4] Melt pool temperature (MPT) [5] Melt pool area (MPA) [6] Load [7] Contact- x_rstc : (N, 50, 24) float32 — radial spatio-temporal context (RSTC) features, 50-step window, standardized. 24 = 3 base RSTC components.- meta : (N, 5) float32 — process-condition fields, stored as raw (unnormalized) values. Columns: [0] layer index [1] Set laser power (W) [2] Set robot speed (m/s) [3] Set wire feed speed (m/min) [4] specimen_id — identifier used for evaluation/analysis only; 0 for non-Overlap experiments, 1-4 for the four specimens per layer in Overlap experiments. Columns [0]-[3] are used as input to the FiLM conditioning path (paper 3.3.2); column [4] is NOT a model input.- y_label : (N,) float32 — working distance label, stored normalized. Apply the scaler inverse transform to recover millimetres (mm). Normalization - x_img and x_rstc are pre-normalized within the .npz files ([0, 1] and standardized, respectively).- x_sensor and y_label are stored as raw values and require the scaler/ parameters (fitted on the training split only) for normalization and for inverse-transforming model outputs to millimetres.- meta is stored as raw values.- Normalization method / library: scikit-learn StandardScaler. Train / validation / test split Splits were created within each experiment as contiguous segments taken from a random start point in the experiment's time series (ratio 8 : 1 : 1). For Overlap experiments, each of the four per-layer specimens was split independently. Splits therefore originate from the same set of experiments/specimens (in-build evaluation) rather than from held-out experiments. Usage Load with numpy.load(file). Each .npz is self-contained (inputs and label per window). The dataset is provided as a single archive (model_input_data.zip) to preserve the folder structure. License Creative Commons Attribution 4.0 International (CC BY 4.0)