Answers are grounded in this paper and cite their sources.
Industrial emission abatement is partially observed: only the stack concentration is measured, whereas pollutant generation and removal efficiency are not. We develop a hybrid soft sensor and a distribution-constrained tandem inverse design and test which of their outputs the data support. On daily data from five sintering units, the test 𝑅 2 of the NO x sensor fell from 0.68 under a random split to 0.18–0.35 under leakage-controlled splits, and SO x was not predictable out of period. The level of the inferred efficiency was not identified, because a ± 20 % bias in the white-box term left accuracy unchanged, but its day-to-day changes were largely identified and tracked the selective catalytic reduction (SCR) inlet temperature, a variable outside the model. For inverse design, self-reconstruction and cross-surrogate 𝑅 2 ranked generated conditions above real ones, whereas realism and coverage metrics reversed this ranking. Prediction intervals lost coverage near the limit under period shift.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: