An accurate reconstruction method for indoor bioaerosol concentration field from asynchronous and sparse LiDAR measurements based on latent diffusion models

• A novel method reconstructs concentration fields from LiDAR scanning measurements. • Latent Diffusion Models enable reconstruction from asynchronous sparse data. • Contrastive Learning links Spatiotemporal sparse measurements to global fields. • Training/testing on CFD-generated data for various bioaerosol release scenarios. • Reconstruction accuracy reaches 80% compared to CFD simulations. Bioaerosols are the primary route for pathogen transmission, and have a serious negative impact on human health. Accurately measuring the distribution of bioaerosol concentrations in the indoor environment holds significant importance. Current methodologies are constrained by limitations in measurement resolution and speed, and unable to obtain the global 3D concentration fields directly. We propose a Sparse Reconstruction Latent Diffusion Model (SR-LDM) that reconstructs the 3D global concentration field from the asynchronous sparse bioaerosol concentration data measured by LiDAR. The concentration field reconstruction is formulated as a conditional generation task, leveraging the powerful generative capabilities of the Latent Diffusion Models (LDM) to infer global distributions from sparse observational data. We employ a contrastive learning strategy to train a dedicated condition encoder, which extracts information from spatiotemporal sparse data measured at different locations and times to guide the generation of the global concentration field. We use a validated Computational Fluid Dynamics (CFD) solver to construct a dataset of aerosol concentration fields in an empty channel. The results indicate that our SR-LDM model effectively reconstructs the global concentration field formed by a single release source, using historical sparse measurement data that accounts for 1%–3.2% of the total number of units. In most cases, high-quality results are obtained after 6-16 iterations, and the similarity with CFD data reaches 80%.