Abstract <p>This study investigates the effect of neural network architecture on the accuracy of data-driven modeling of thermal explosions in a hydrogen-oxygen-air mixture. Using a reduced kinetic mechanism for 11 reagents, the thermal explosion process is simulated under specified initial pressure and temperature conditions, generating time-resolved data. We compare three architectures: a standard multilayer perceptron (MLP), a DeepONet-inspired model, and our U-Net-style residual network, evaluating their ability to capture transient dynamics and key reaction regimes. Our results demonstrate that network architecture has an important impact on predictive performance. Among the three considered architectures, the U-Net-based model demonstrated the lowest prediction error, with mean squared error (MSE) of 1.3 × 10^–3 and a standard deviation (STD) of 2.18 × 10^–2, indicating stable performance across the test set. By comparison, both the DeepONet-inspired model and the MLP exhibited substantially larger errors, with MSE values of 1.81 × 10^–2 (STD 5.81 × 10^–2) and 2.02 × 10^–2 (STD 6.82 × 10^–2), respectively. The increased dispersion of errors for these two models was particularly pronounced in trajectories involving rapid temperature growth. These findings suggest that the careful choice of network architecture is a critical factor in the development of reliable neural-network surrogates for combustion and reactive-flow modeling.</p>