Time-aware conformal uncertainty quantification with physics-consistent projection for fusion blanket activation time-series prediction

Abstract

Predicting radioactivity-related time series in fusion blankets requires modelling targets spanning multiple orders of magnitude under nonnegativity and smoothness constraints, where standard conformal prediction (CP) yields miscalibrated intervals because residuals are non-stationary across physical regimes and post-hoc projections enforcing physical limits invalidate coverage guarantees. We present a physics-constrained CP framework with three components: time- and phase-aware split conformal inference conditioning residual quantiles on discrete operational time bins; a convex projection operator bounding second differences and enforcing nonnegativity; and a projection-aware calibration loop scoring against post-projection estimates. A geometry-level score aggregation strategy changes the controlled event from point-wise marginal to bin-wise simultaneous coverage: for each temporal bin, all time points within the bin are simultaneously covered for a new test geometry with probability ≥ 1-α. Evaluated on coupled neutronics–activation datasets (224 PbLi blanket geometries, 15 targets), a cross-conformal extension with half-life-adaptive projection (SC-PIML+) attains 15/15 targets on the a priori fixed canonical partition under point-wise marginal coverage, outperforming all five baselines, with raw PICP ≥95% on every target and bootstrap-significant coverage on 9 of 15; the framework is architecture-compatible and validated across seven base learners spanning four algorithmic families, with tree-based models giving the best coverage–efficiency tradeoff. Under a stricter 90% bin-wise simultaneous criterion across 20 random partitions the method passes of 4 representative targets; that criterion sits above the ≈86% finite-sample level of the geometry-level CV+ construction (whose split-conformal variant carries ≥ 1-α), so the two studies are reported separately and are not directly comparable estimates of the same performance metric. An operational out-of-distribution detection metric, validated for composition-shift scenarios, enables automated deployment rejection, and the prediction intervals quantify surrogate model-form uncertainty relative to the FISPACT-II reference solver.