Carbon intelligence for industrial decarbonisation: a physics-anchored machine learning architecture for emission-intensity forecasting

Cement manufacture is among the most emission-intensive industrial activities, and data-driven carbon accounting is increasingly proposed to supplement conventional activity-factor inventories. This work examines a question that the carbon-analytics literature has largely left implicit: what happens to a learned emission model when the plant it monitors is itself decarbonising. A physically consistent hourly digital twin of a dry-process precalciner cement plant is constructed from mass- and energy-balance relations, spanning three years and 24,698 operating hours, with a mean Scope 1 and Scope 2 intensity of 630.0 kg CO2e per tonne of cement. Five learning algorithms, namely ridge regression, support-vector regression, random forest, gradient boosting and a multilayer perceptron, are benchmarked on twenty-four-hour-ahead intensity forecasting under a strictly chronological protocol, alongside a closed-form physics baseline and a proposed physics-anchored residual architecture designated CIRCE-Net. Two results stand out. First, every locally supported learner exhibits a positive test bias, ranging from 2.4 to 16.3 kg CO2e per tonne, because a falling target cannot be extrapolated from historical support; regularised linear regression is the only method with negative bias and attains the lowest test error of 5.241 kg CO2e per tonne. Second, the ranking obtained on an adjacent validation year does not reproduce on the deployment year. CIRCE-Net reduces the gradient-boosting error by 23.1 per cent and is the strongest method under a deliberate high-substitution stress test, where it attains 4.705 kg CO2e per tonne with near-zero bias, yet it remains significantly inferior to ridge regression on the standard test year. Ablation shows that intensity history is actively harmful and that weather forecasts contribute nothing. Unsupervised multivariate detectors operate at chance for excursion detection, while a residual-based monitor reaches an area under the curve of 0.743 with strongly fault-dependent recall. Negative and non-confirmatory findings are reported in full.