SBS Nanofiber Diameter Prediction: Dataset, Trained Models, and Results

This archive is the frozen, citable companion to a study on predicting solution blow spinning (SBS) nanofiber diameter from physics-informed features using machine learning and deep learning. It preserves everything needed to reproduce the reported results: the complete dataset, every trained model, all fitted scalers, the external-laboratory validation data, and the full set of result tables and serialized analysis objects. STUDY AT A GLANCE- Task: regression of fiber diameter, modeled as natural-log diameter in nm.- Dataset: 408 samples drawn from 57 published studies covering 29 polymer systems.- Features: eight physics-informed inputs — polymer concentration (wt%), intrinsic viscosity (dL/g), reduced concentration c/c*, solvent surface tension (mN/m), solvent boiling point (degC), standardized air pressure (bar), standardized feed rate (mL/h), and working distance (cm).- Models: eight machine-learning models (ExtraTrees, RandomForest, XGBoost, LightGBM, CatBoost, SVR, ElasticNet, KNeighbors) and four deep-learning models (MLP, 1D CNN, TabNet, FT-Transformer).- Validation: three levels — (1) grouped cross-validation, (2) a held-out set of unseen studies for zero-shot and calibrated evaluation, and (3) an independent external laboratory validation set, with post-hoc affine calibration for cross-laboratory transfer. CONTENTS- dataset/ — the full extracted dataset, the model-ready table (features plus log-diameter target), the raw pure-polymer records, the cross-validation and held-out splits, the external validation samples, the split definition (split_info.json), and solvent property lookups.- models/ — the eight trained machine-learning models and four deep-learning models, the fitted scalers, and a portable deep-learning export (PyTorch state dictionaries, TabNet model, architecture configuration, and model-class definitions).- external_validation/ — the external-laboratory feature table and 21 scanning electron micrographs (S01–S21) of the externally spun samples, including defect and no-fiber cases.- results/ — result tables as CSV files (dataset summary, tuned hyperparameters, held-out and external performance, per-study calibration, benchmark comparison, feature ablation, statistical tests, SHAP feature importance, per-study statistics, and the intraclass-correlation analysis) and the serialized objects backing the tables and figures. LICENSEThe data, models, and results in this archive are released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. The analysis source code in the companion GitHub repository is released under the MIT license. RELATED RESOURCESAnalysis source code (GitHub): https://github.com/Mohammadlari977/sbs-fiber-predictionInteractive prediction tool, SpinLab (Hugging Face): https://huggingface.co/spaces/mohammadlari97/sbs-fiber-predictor