Graph Neural Operator-Based Surrogate Modelling of Multi-Field CFD Results in Biomass Boiler

Computational fluid dynamics provides detailed spatial distributions of physical fields in biomass boiler combustion, but the computational cost of each simulation limits its application in parametric studies and near-real-time workflows. This work investigates whether a Graph Neural Operator (GNO) can serve as a fast surrogate model that maps boiler operating parameters to six coupled CFD field distributions simultaneously. The reference case is a 10 kW wood-pellet boiler with internal flue gas recirculation (FGR), described and experimentally validated in an earlier publication by the authors. CFD data were collected on the symmetry plane of the combustion chamber for 80 operating points defined by the thermal load ratio (P/P0) and the excess air ratio λ. A GNO surrogate was trained on 64 cases to predict temperature, velocity magnitude, static pressure, and mole fractions of CO, O2, and CO2 at each node of an unstructured spatial graph. On a held-out validation set of 16 operating cases, the model achieved R2 values of 0.988 for temperature, 0.919 for velocity magnitude, 0.982 for pressure, 0.999 for CO, 0.992 for O2, and 0.985 for CO2. After training, each prediction is generated in a single forward pass, providing a computationally efficient approximation compared to the full CFD solver. A dedicated generalisation study on independent off-grid CFD cases confirmed that the surrogate interpolates within the parameter domain with essentially no loss of accuracy and degrades only moderately when extrapolated towards a higher thermal load and leaner mixtures. The results demonstrate that a baseline GNO surrogate can capture the spatial structure of coupled thermo-fluid and species fields in a realistic combustion geometry within the investigated parameter range and suggest applicability to digital-twin-oriented workflows where repeated parametric queries of boiler operation are required.