In pursuit of a low-carbon energy transition, biomass and other carbon-neutral fuels are increasingly utilized in modern combustion facilities. However, controlling these systems remains challenging due to their complex geometries, dynamic interactions, and diverse operating conditions. Data-driven digital twins have emerged as powerful tools for optimizing performance and minimizing emissions in industrial combustion systems. Their core functions include reconstructing multi-physical combustion fields and predicting and optimizing key performance metrics, such as efficiency and pollutant emissions. Despite recent advancements, existing approaches typically treat reconstruction and optimization as separate tasks, limiting their efficiency and scalability. Furthermore, developing digital twins for real-world industrial applications requires extensive high-fidelity data, which is often impractical to obtain. To address these limitations, we propose the multi-field reconstruction net (MFRNet) framework, which integrates dimension expansion, variable extension, and feature fusion techniques to enhance data efficiency and predictive accuracy. Using an industrial-scale biomass grate furnace as a case study, we construct a comprehensive dataset, consisting of 288 low-fidelity 2D cases (covering eight physical fields) and 48 high-fidelity 3D cases (covering eleven physical fields). The MFRNet achieves high-precision multi-field reconstruction under complex conditions while significantly reducing the reliance on costly 3D simulations. By leveraging intermediate features pre-trained during reconstruction, the model enhances scalar predictions, notably improving CO and NO emission accuracy, even with limited high-fidelity data. The trained model is then directly applied for multi-objective optimization under varying operating conditions, demonstrating robust predictive accuracy and reliable optimization guidance. This scalable and data-efficient digital twin framework is easily adapted for other combustion systems, offering an intelligent paradigm for active control, real-time optimization, and enhanced operational efficiency in modern combustion facilities.