SC-PIML+: Time-aware conformal uncertainty quantification with physics-consistent projection — dataset and reproducibility package

This Zenodo archive accompanies the manuscript "Time-aware conformal uncertainty quantification with physics-consistent projection for radioactivity-related time-series prediction" by Xiaokang Zhang, Xilong Tong, Yanshi Wei, and Shanliang Zheng (corresponding author), submitted to Machine Learning: Science and Technology (MLST), April 2026. Scientific scope. Predicting radioactivity-related time series in fusion blankets — activation inventories, decay heat, and contact dose — requires modeling targets spanning multiple orders of magnitude while respecting physics-based nonnegativity and smoothness constraints. This dataset and reproducibility package supports a physics-constrained conformal prediction (CP) framework with three components: (i) time- and phase-aware split conformal inference that conditions residual quantiles on discrete operational time bins; (ii) a convex projection operator bounding second differences and enforcing nonnegativity; and (iii) a projection-aware calibration loop that scores against post-projection estimates to preserve coverage. Contents. D224+act9 dataset — coupled neutronics–activation data for 224 PbLi blanket geometries × 9 activation/decay-heat targets, generated with the FISPACT-II v5.0 depletion solver (UKAEA). Train / calibration / test splits used in the paper, including the canonical D224+act9 split and the 20 random-partition replicates. SC-PIML+ calibration run outputs — point predictions and 90% conformal intervals for all six benchmark methods (Split CP, CQR, ACI, EnbPI, SC-PIML, SC-PIML+). Per-time-step coverage and width metrics across all 15 targets, supporting reproduction of all reported PICP and Winkler interval-score figures. Out-of-distribution validation batch (D70-valid, N = 70) with shifted geometry, neutron spectra, and impurity compositions, used to demonstrate the operational OOD detection gate. Figure-generation scripts sufficient to reproduce all figures and tables in the manuscript. Versioning. The concept DOI 10.5281/zenodo.19679269 always resolves to the latest version. The version-specific DOI for this release is 10.5281/zenodo.19679660. The corresponding GitHub release tag is v1.0.3-mlst-submission. License. Released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). The FISPACT-II solver used to generate the underlying activation data is developed and licensed by UKAEA and is not redistributed here.