Speaker
Description
Machine-learning surrogates for coupled neutronics–activation calculations (OpenMC + FISPACT-II) reduce per-sample evaluation time from ~40 min to below one second, making rapid blanket design iteration feasible. Point predictions, however, do not suffice for safety-critical decisions. Activation inventory, decay heat, and contact dose rate span more than six orders of magnitude across time scales from hours to $10^5$ years; credible uncertainty intervals are therefore indispensable.
Standard conformal prediction (CP) produces miscalibrated intervals in this setting because (i) predictive residuals are non-stationary across irradiation and cooling regimes, and (ii) physical constraints—nonnegativity and monotonic decay—invalidate coverage guarantees when enforced as post-hoc corrections. We present SC-PIML+, a physics-constrained conformal framework comprising: (1) physics-informed feature engineering with capped half-life detrending; (2) time- and phase-stratified nonconformity scoring; (3) a convex projection operator preserving nonnegativity and smoothness in log-space, applied before calibration scoring to maintain coverage validity; and (4) a cross-conformal (CV+) extension with half-life-adaptive projection policy that increases effective calibration sample size by ~30×.
We evaluate SC-PIML+ on 224 CFETR PbLi blanket geometries with 15 prediction targets in four categories: breeding isotopes ($^3$H, $^6$Li, $^7$Li), aggregate safety indicators (decay heat, total activity, contact dose rate), PbLi-origin activation products ($^{210}$Po, $^{210}$Bi, $^{203}$Hg, $^{204}$Tl, $^{207}$Bi, $^{203}$Pb), and steel-origin activation products ($^{54}$Mn, $^{60}$Co, $^{55}$Fe). In a six-method benchmark, SC-PIML+ is the only method that meets nominal 95% prediction interval coverage on all 15 targets; Split Conformal covers 12, Conformalized Quantile Regression 11, and the base SC-PIML 13. Across 20 independent random data partitions the method passes 10.5 ± 2.4 targets on average; the remaining spread stems from finite-sample calibration ($n_\text{cal} \approx 30$), not from method instability. Because the conformal calibration is decoupled from the base learner, gradient-boosted trees, MLPs, and LSTMs all run through the same pipeline without modification. An operational out-of-distribution detection metric flags extrapolation cases for automatic fallback to direct FISPACT-II evaluation.
| Formatted abstract uploaded? | Done. |
|---|