Multimodal SEM Dataset of Non‑metallic Inclusions in Ti-ULC Steel: BSE Images and EDS Elemental Compositions

This dataset contains backscattered electron images and quantified energy-dispersive X-ray spectroscopy (EDS) compositions of features detected by automated SEM/EDS analysis of Ti-ULC steel samples, together with manually verified binary labels indicating whether each feature is a true or a false non-metallic inclusions (artifact misclassified as an inclusion by compositional filtering). Contents features_train_public.tar — training split features_val_public.tar — validation split features_test_public.tar — test split metadata_public.json — class counts, EDS standardization statistics, and split sizes File format The dataset is stored in the WebDataset format. Each sample consists of three files sharing the same key prefix: image.png — local BSE image patch around the feature, 16-bit grayscale, variable size, pixel size 0.33 µm eds.json — quantified EDS composition in wt%, dictionary of element symbol to weight percentage (17 elements) label.cls — binary label, 1 for true NMI, 0 for false NMI Loading example python import webdataset as wdsfrom PIL import Imageimport ioimport jsonimport numpy as np dataset = wds.WebDataset("features_train_public.tar") for sample in dataset: img = np.array(Image.open(io.BytesIO(sample["image.png"]))) eds = json.loads(sample["eds.json"]) label = int(sample["label.cls"]) # img: (H, W) uint16 array # eds: {"C": 5.51, "O": 18.38, ..., "Fe": 53.88}, values in wt% # label: 1 = true NMI, 0 = false NMI break Data acquisition Features were detected by automated SEM/EDS analysis of polished Ti-ULC steel samples using a JEOL JSM-7200F field emission scanning electron microscope equipped with an Oxford Instruments Ultim Max 100 silicon drift detector. Acquisition parameters were 15 kV acceleration voltage, 400× magnification, 0.33 µm/pixel image resolution, and 1 to 3 s EDS analysis time per feature. Features with an equivalent circle diameter below 2 µm were excluded to ensure sufficient morphological resolution in the BSE image. Ground truth labels were assigned by manual verification of each feature using both BSE morphology and EDS composition. Preprocessing notes EDS values are provided as raw weight percentages. The metadata_public.json file contains the per-element mean and standard deviation computed on the training set, which can be applied to standardize the EDS vector before model training. See the accompanying publication for the preprocessing details.