Ultra-high performance concrete (UHPC) combines exceptional strength and durability, yet its industrial production is hampered by batch-to-batch variability that generates costly off-specification waste. Leveraging a 150-batch design-of-experiments dataset based on systematic variations of a single reference UHPC mix, this study takes a holistic view of the UHPC manufacturing chain and quantifies how fluctuations in raw material quality, storage conditions, dosing errors, mixer energy demand, and curing regimes affect the 28-day compressive strength. Ten diverse machine learning algorithms are benchmarked; the best-performing model explains \(\ge\) 75 % of the strength variance with a prediction error \(\le\) 10 % under leave-one-out cross-validation. SHapley Additive exPlanations reveal that long-term curing temperature and humidity dominate strength development, followed by ingredient moisture and silica fume impurity. These insights are operationalized in an at-line, operator-in-the-loop recommendation system that explores the curing envelope and proposes end-of-mix, batch-specific adjustments before curing starts. In five validation cases, curing adjustments rescued 5/5 underperforming batches, eliminating 75 L of off-specification UHPC and—considering cement only with 600 kg/ \(\mathrm {m^{3}}\) and 15 L per batch of UHPC made with white Portland cement—avoided \(\approx\) 41 kg CO2e (cement-only; 0.913 kg CO2e/kg, A1–A3). The framework therefore not only elucidates the main sources of UHPC quality inconsistency but also provides a practical, data-driven tool to rescue off-specification products, minimize waste, and cut associated \(\mathrm {CO_2}\) emissions.
