Ex-Ante Product Carbon Footprint Forecasting in Electric Arc Furnace Stainless Steel Production: A Probabilistic Machine Learning Framework for Alloy Eco-design

Life Cycle Assessment (LCA) and Product Carbon Footprint (CFP) are robust tools for evaluating sustainability in hard-to-abate sectors like steelmaking, but their data-intensive and ex-post nature severely limits their use as proactive eco-design tools. Highly granular LCA models mapping entire product portfolios remain exceptionally rare, as their development is labor-intensive and demands specialized expertise. To overcome this bottleneck and tackle the eco-design paradox, this work proposes an ex-ante Machine Learning (ML) framework to forecast the CFP of Electric Arc Furnace (EAF) stainless-steel production, directly during the alloy and charge design stage. The ML framework was developed by analyzing 2206 EAF heats. The target CFP for each heat was derived from a previous study and charge mix macro-categories (i.e., scrap, Cr, Ni, and Mo sources) were used as features. This study overcomes deterministic ML approaches by propagating the LCA model uncertainty via dataset augmentation, leveraging Monte Carlo simulations with 10,000 iterations per steel family. Algorithms including Linear Quantile Regression (LQR), XGBoost, Quantile Random Forest (QRF), Multilayer Perceptron (MLP), Long Short-Term Memory neural networks (LSTM) and TabNet were evaluated. To ensure methodological robustness and industrial reliability, a time-series cross-validation (walk-forward) approach was employed to prevent look-ahead bias and data leakage, effectively handling temporal variations in process conditions (e.g., maintenance, refractory conditions) and scrap availability. SHAP (SHapley Additive exPlanations) analysis was used to interpret feature contributions, acting as a consistency check. QRF emerged as the most robust algorithm, outperforming other models due to its generalization capabilities and its ability to capture LCA parametric uncertainty. The model achieved an R2 of 97.4% and a Mean Absolute Error (MAE) of 105.4 kg CO2,eq/ton (a 4.38% relative error against the full dataset average of 2404.65 kg CO2,eq/ton). Furthermore, the model successfully provided probabilistically calibrated prediction intervals that robustly account for the uncertainty stemming from the underlying LCA model, achieving a Prediction Interval Coverage Probability (PICP) of 95.3%. SHAP analysis confirmed the model's reliance on raw materials, accurately capturing the strong positive correlation between CFP and main alloying sources (i.e., Cr, Ni, and Mo). Conversely, a negative correlation was found with scrap. By linking raw material inputs to environmental performance, the proposed ML framework provides a robust, transferable, and scalable eco-design tool. This empowers steelmakers to optimize secondary raw material usage and improve alloy design, thereby accelerating the transition towards resource-efficient and sustainable metallurgy.