<p>Well-estimated air pollutant concentration fields through data fusion are critically important to compensate the observations that are only sparsely available, especially over non-urban areas. Previous data fusion methods generally used statistical models to relate target observations and supporting data variables at known stations. In this study, we built a new data fusion paradigm by designing a dedicated deep learning framework to learn multi-variable spatial correlations from Chemical Transport Model (CTM) simulations, before using it to estimate PM<sub>2.5</sub> reanalysis fields from station observations. The model was composed of two modules, which include an explainable PointConv operation to pre-process isolated observations and a regression grid-to-grid network to reflect correlations among multiple variables. The model was evaluated in two aspects of reproducing PM<sub>2.5</sub> CTM simulations and generating reanalysis/fused PM<sub>2.5</sub> fields. First, the fusion model was able to well reproduce CTM simulations from sampled station CTM data items with an average R<sup>2</sup> = 0.94. Second, the fusion model achieved good performance with R<sup>2</sup> = 0.77 and R<sup>2</sup> = 0.83 respectively evaluated at the stringent city-level and station-level. The generated reanalysis PM<sub>2.5</sub> fields have complete spatial coverage within the modelling domain and at daily time scale. One significant benefit of our fusion framework is that the model training does not rely on observations, which can be used to predict PM<sub>2.5</sub> fields in newly-setup observation networks such as those using portable sensors. The fusion model has high computing efficiency (< 1 s/day) in predicting PM<sub>2.5</sub> concentrations due to acceleration using GPU. As an alternative to generate chemical/meteorological reanalysis fields, the method can be readily applied for other simulated variables that with measurements available.</p>