Accelerating life cycle inventory development through correlation-aware multi-fidelity transfer learning: Case studies in cradle-to-gate manufacturing energy modeling

Accurate, parametric cradle-to-gate Life Cycle Assessment (LCA) models of production systems are essential for environmentally motivated process comparison and optimization; however, commercially-available Life Cycle Inventory (LCI) databases are often too inaccurate for confident decision-making, while more accurate process and/or site-specific data-driven modeling can require prohibitively expensive experimental campaigns. In this work, a multi-fidelity transfer learning (MFTL) framework is developed for cradle-to-gate LCA that combines a low-fidelity reduced-order model, embedding basic physics or experience as a principled head start, with a high-fidelity adaptive boosting mechanism based on limited experimental data. The framework accelerates learning and helps prevent overfitting by augmenting the traditional residual-based loss with a correlation-aware term. The framework is demonstrated by modeling energy-related inventory data in two manufacturing processes: roller power consumption in aluminum cold rolling and energy use in polymer extrusion-based additive manufacturing (AM). The correlation-aware MFTL framework achieved R 2 validation scores of 0.99 and 0.98 using just 60% and 12% of the data collected from experiments on the rolling and AM case studies, respectively. This predictive accuracy outperformed Traditional MFTL ( R 2 validation scores of 0.88 and 0.83) and purely data-driven approaches such as a neural network and a support vector regressor ( R 2 validation scores of 0.85 and 0.84) when trained on the same amount of data. Overall, the framework enables accelerated, accurate, parametric modeling of LCI data, supporting the development of sustainable manufacturing systems.