AI-Guided Predictive Carbonation Mapping for Ultra-Low-Emission Cement Manufacturing Facilities

Abstract

Chinenye Elizabeth Onumadu

Carbonation curing can transform cement manufacturing from a CO2 source to a CO2 sink, but its industrial adoption is limited by unpredictable outcomes arising from feedstock variability, fluctuating flue gas compositions, and inhomogeneous chamber conditions. Current control strategies rely on static lookup tables that fail to account for real- time process dynamics, leading to underperformance or batch rejection. Here we develop and validate an AI-guided predictive carbonation mapping system that forecasts four key performance indicators simultaneously: CO2 absorbed (wt%), carbonation depth (mm), 28-day compressive strength (MPa), and porosity reduction (%). Using a dataset of 10,800 production batches from a pilot carbonation chamber operating over 12 months (24 input features including raw meal XRF, chamber temperature, relative humidity, pCO2, and curing duration), we trained a hybrid architecture combining a physics-informed neural network (PINN) embedding a simplified diffusion-reaction PDE with a gradient- boosted residual corrector (XGBoost). The hybrid model achieves hold-out test R2 values of 0.94 (CO2 uptake), 0.91 (strength), 0.89 (carbonation depth), and 0.87 (porosity reduction), significantly outperforming standalone machine learning baselines (best baseline R2 = 0.85). SHAP analysis reveals pCO2 (33% contribution), curing temperature (22%), and water-to-cement ratio (18%) as dominant predictors, with reactive MgO content (9%) unexpectedly influential. The model generates 2D spatial carbonation maps revealing heterogeneous uptake (up to 20% variation across chamber corners), enabling targeted process redesign. Real-time inference runs at 47 ms per batch with uncertainty bounds (90% CI width ±1.8% CO2 uptake), sufficient for closed-loop control. Simulated deployment reduces batch failure rates from 18% (conventional lookup-table control) to 6% while increasing average CO2 uptake by 15% – equivalent to 75 kt additional CO2 sequestered annually per 0.5 Mt plant. Key limitations include single-plant training chemistry and batch-averaged inputs; future work requires transfer learning across geologies and per-sample near-infrared sensors. This work demonstrates that hybrid AI, combining physical constraints with data-driven flexibility, can transform carbonation curing from an art to a predictable, optimizable industrial process.

PDF