Dimensionality reduction of atmospheric concentration fields using a 3D convolutional autoencoder with spatial attention mechanism

Abstract High-dimensional atmospheric chemistry simulations produce volumetric concentration fields that are costly to store and repeatedly process in inverse-modelling and data-assimilation workflows. This study evaluates reduced representations for CO concentration fields from WRF-Chem over Novosibirsk and CAMS over Europe. We propose a three-dimensional convolutional autoencoder with a lightweight spatial attention module (SAM3D) and compare it with linear, spectral, tensor-based, interpolation, manifold-learning, and neural-network baselines. The evaluation includes latent dimensions from 8 to 64, temporally separated validation protocols, storage-cost estimates, and computational-efficiency measurements. On the larger CAMS dataset, SAM3D achieves the lowest MSE and relative $$L_2$$ error among methods compared at equal latent-vector dimensions and improves over the plain 3D autoencoder. On the smaller WRF-Chem dataset, rankings depend more strongly on the split protocol, with fixed transforms and tensor decompositions remaining competitive. SAM3D adds less than 1% FLOPs and only a small increase in model size while retaining a differentiable encoder–decoder mapping. These results support SAM3D as a practical reduced-state representation for three-dimensional atmospheric concentration fields.