<p>Accurate prediction of lime and limestone consumption in converter steelmaking is essential for optimizing steelmaking efficiency and ensuring coordinated control of endpoint temperature and slag basicity. To address this challenge, a dual-stream network, characterized by parallel feature extraction, cross-stream information interaction, and mechanism-embedded loss function design, is proposed to improve prediction accuracy and robustness of the coupled multi-objective optimization problem. Unlike conventional approaches that treat lime and limestone prediction as independent tasks or suffer from error propagation in sequential modeling, the proposed method effectively captures the strong coupling relationship between temperature control and slag chemistry through incorporating metallurgical principles directly into the optimization objective, enabling more accurate, stable, and physically consistent predictions in converter steelmaking processes. Extensive experiments conducted on 5331 industrial converter operation samples demonstrate that the proposed model achieves a mean of 0.46, significantly outperforming chain prediction methods and independent modeling approaches, with prediction accuracy improvements of 2–4 times and 35–46 times, respectively. These results highlight the potential of the proposed dual-stream network to provide reliable decision support for converter process control through simultaneous optimization of temperature–basicity objectives.</p>