Traditional computational tomography methods entail substantial computational costs for 3D reconstruction and time series prediction of flames in combustion fields, making real-time prediction of projection information challenging. The advancement of deep learning has facilitated 3D reconstruction, enhanced data processing efficiency, reduced experimental costs, and enabled real-time prediction capabilities. The computational efficiency and accuracy of combustion length prediction can be enhanced through the development of more efficient networks, optimization of model parameters, and improvements in computing capabilities.

The GAST-Net employs convolutional neural networks to capture complex spatial structural features, integrating gated recurrent units (GRUs) to address long-distance dependencies. The time processing module is integrated into the network’s skip connections to enhance temporal information modeling through spatiotemporal feature fusion. These skip connections mitigate the gradient vanishing problem during deep network training. Additionally, a hybrid attention mechanism in the time processing module separately processes critical temporal and spatial information at each feature layer, achieving comprehensive information fusion and utilization while significantly improving prediction and reconstruction accuracy. A projection acquisition device based on computed tomography of chemistry was constructed, and the three-dimensional structure distribution of CH^* radicals in the combustion field was reconstructed using traditional methods as a ground truth reference for neural network training.

Initial validation of GAST-Net demonstrated its capability to predict reconstruction structure from historical multi-directional projection data within 7 ms (Fig. 4), maintaining high prediction reconstruction accuracy across multiple time intervals (Fig. 5). Further analysis comparing predicted 3D field structure data with 2D slices of real reconstruction data confirmed accurate prediction of internal combustion field structures. Quantitative evaluation through four quality assessment functions revealed that prediction accuracy decreases with increasing time steps while maintaining high performance levels. The model demonstrated effectiveness in inter-frame prediction of 3D fields (Fig. 8) and maintained robust performance when multiple instantaneous 3D data were inserted between consecutive time series (Fig. 9). Both future time prediction and inter-frame prediction scenarios achieved structural similarity exceeding 0.93, correlation coefficients above 0.9, root mean square error below 0.02, and peak signal-to-noise ratio not less than 34 (Figs. 7 and 10). Additionally, ablation experiments examined the impact of attention module placement and various network parameters on reconstruction accuracy.

This study presents a novel emission spectrotomography reconstruction prediction method incorporating gated recirculating units and attention mechanisms. The approach utilizes traditional chemiluminescence computational tomography algorithm reconstructions of CH^* radical three-dimensional distributions as reference data for deep learning validation. GAST-Net leverages gated recurrent units for managing long-distance dependencies and attention mechanisms for importance weight adjustment, enabling three-dimensional reconstruction and future prediction of combustion fields from historical multi-directional two-dimensional projections. This methodology eliminates the need for complex acquisition devices or partial differential equation solutions. Experimental results demonstrate rapid and accurate prediction of future combustion evolution based on continuous historical multi-directional projection data. Furthermore, the model successfully addresses interpolation between continuous time series, accurately predicting multi-frame 3D structures using before-and-after multi-directional projections. These findings indicate significant potential for addressing time series challenges and suggest new approaches for solving long-distance dependency problems. The network’s prediction accuracy and applicability can be further enhanced through expanded training data.