Integrated Experimental and Physics-Informed Neural Networks Assessment of Emissions from Pelleted Woody Biomass

Accurately predicting pollutant emission factors (EFs) from woody biomass fuels remains challenging because small-scale combustion tests are fuel-specific, time-consuming, and highly sensitive to operating conditions. This study combines controlled laboratory combustion experiments with a physics-informed artificial neural network (ANN–PINN) to estimate the emission factors of particulate matter (EFPM), carbon monoxide (EFCO), and nitrogen oxides (EFNOx) using only laboratory-scale fuel characterization. Three pelletized woody biomass, Pinus radiata, Acacia dealbata, and Nothofagus obliqua, were analyzed through ultimate and proximate composition, lignin content, and TGA-derived parameters and tested in a residential pellet stove under identical control setpoints, resulting in a narrow and well-defined operating regime. A medium-depth ANN–PINN was constructed by integrating mechanistic constraints, monotonicity based on known emission trends and a weak carbon balance penalty, into a feed-forward neural network trained and evaluated using Leave-One-Out Cross-Validation. The model accurately reproduced the experimental behavior of EFCO and captured structured variability in EFPM, while the limited nitrogen variability of the fuels restricted generalization for EFNOx. Sensitivities derived via automatic differentiation revealed physically coherent relationships, demonstrating that PM emissions depend jointly on fuel chemistry and aero-thermal conditions, CO emissions are dominated by mixing and temperature, and NOx formation is primarily governed by fuel-bound nitrogen. When applied to external biomass fuels characterized independently in the literature, the ANN–PINN produced physically plausible predictions, highlighting its potential as a rapid, low-cost screening tool for assessing new biomass feedstocks and supporting cleaner residential heating technologies. The integrated experimental–PINN framework provides a physically consistent and data-efficient alternative to classical empirical correlations and purely data-driven ANN models.