A research team from the Massachusetts Institute of Technology has introduced a data-driven digital twin framework named Multi-Field Reconstruction Net (MFRNet) for industrial-scale combustion systems, aiming to support low-carbon energy transition with improved data efficiency and multi-task integration. The work was published in Engineering, providing a scalable solution for 3D multi-physical field reconstruction, emission prediction, and operational optimization under limited high-fidelity data conditions.

Against the background of growing use of carbon-neutral fuels such as biomass in combustion facilities, researchers note that conventional digital twin approaches often separate reconstruction and optimization processes, and demand extensive high-fidelity 3D simulation data that is computationally costly to generate. The MFRNet framework integrates dimension expansion, variable extension, and dynamic feature fusion to address these limitations, unifying multi-field reconstruction and multi-objective optimization within one machine learning pipeline.

In the validation using an industrial biomass grate furnace, the team constructed a hybrid dataset including 288 low-fidelity 2D cases covering eight physical fields and 48 high-fidelity 3D cases covering eleven physical fields. The model is first pre-trained on low-fidelity 2D data to explore the high-dimensional condition space, then fine-tuned on a small set of 3D data through dimension expansion to capture z-direction heterogeneity and boundary effects. Variable extension modules further infer NOₓ-related species distributions including NO, HCN, and NH₃ by reusing latent features from core combustion variables learned during pre-training.

The study shows that MFRNet supports multi-modal inputs consisting of operational parameters and sparse temperature measurements, aligned via contrastive learning to enable robust reconstruction even with partial input information. By leveraging intermediate features from the reconstruction stage, the framework enhances scalar prediction for key indicators such as CO and NO emissions at the furnace outlet. The fused features improve prediction accuracy compared with direct neural network mapping, supporting reliable response surface construction.

The trained digital twin is applied to multi-objective optimization using the NSGA‑II algorithm, targeting minimized CO and NO emissions under fixed capacity conditions. The optimization generates Pareto fronts that reveal trade-offs between combustion efficiency and pollutant control, supporting the identification of practical operating strategies. The framework maintains high precision in 3D multi-field reconstruction while substantially reducing dependence on computationally expensive 3D simulations, showing adaptability to diverse industrial combustion systems.

This integrated digital twin approach offers a data-efficient pathway for active control and real-time optimization of modern combustion facilities, supporting stable operation and low-emission performance in the shift toward low-carbon energy systems.

The paper “Building Digital Twin for 3D Multi-Field Reconstruction and Optimization of Industrial-Scale Combustion Systems,” is authored by Linzheng Wang, Yaojun Li, Sili Deng. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.08.020. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.