Abstract
<title>Abstract</title> <p>High-dimensional atmospheric chemistry simulations generate volumetric concentration fields that are costly to store and repeatedly process. Compact representations are particularly relevant as a preliminary stage in inverse-modelling and data-assimilation workflows, where full-resolution fields may be processed many times. This study evaluates dimensionality-reduction methods for carbon monoxide fields produced by WRF-Chem model over the Novosibirsk urban area. We propose a three-dimensional convolutional autoencoder with a lightweight spatial attention module(SAM3D) designed to preserve localized sources and sharp concentration gradients under strong compression. The model is compared with linear, spectral, tensor-based, interpolation, manifold-learning, and neural-network approaches with latent space dimensions from 8 to 64. SAM3D consistently improves reconstruction over the plain autoencoder, reducing mean squared error by up to 37%, and outperforms the main linear baselines at latent dimensions of 16, 32, and 64. These results establish a basis for future integration into inverse-modelling workflows.</p>