Laminar Burning Velocity of Ammonia‐Hydrogen Mixtures Predicted by Machine Learning

ABSTRACT

Ammonia–hydrogen blends are vital for global defossilization, combining hydrogen's exceptional gravimetric energy density with ammonia's superior storage advantages. However, pure ammonia exhibits poor chemical kinetics and a slow laminar burning velocity (LBV), necessitating precise modeling. Because chemical kinetic modeling is computationally prohibitive, machine learning (ML) surrogates offer a much faster alternative. This study evaluated the performance of three standard ML algorithms—Gaussian process regression (GPR), decision trees (DT), and artificial neural networks (ANN)—using 1445 real experimental data points under data‐scarce conditions. Under systematic dataset reduction, a clear algorithmic hierarchy emerged. GPR demonstrated superior robustness, maintaining highly stable predictive accuracy down to a lower bound of approximately 100 data points. Conversely, the DT model's capabilities rapidly deteriorated below 700 data points, whereas the ANN exhibited a strong propensity for overfitting, making GPR the most reliable choice.