Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper evaluates the performance of various deep generative models (DGMs) in reproducing non-stationary Gaussian random fields, a crucial aspect for modeling complex spatial and spatio-temporal data. The authors assess the models—flow matching (FM), DDPM, score-SDE, and VAE—using comprehensive metrics that measure the recovery of ground-truth mean and covariance structures against oracle samples and a stationary control. While all models successfully recover the mean surface, their effectiveness in capturing covariance varies significantly, with DDPM and score-SDE performing well, FM showing mild attenuation, and VAE struggling with covariance recovery.
Covariance recovery in deep generative models varies dramatically, with DDPM and score-SDE outperforming FM and VAE in capturing non-stationary structures.
Deep generative models (DGMs) are widely used for complex high-dimensional data and increasingly applied to spatial and spatio-temporal modeling. Their generated samples implicitly represent the learned data distribution and associated uncertainty. However, for real-world data, assessing whether DGMs have learned the underlying process is difficult because the ground truth is unknown and evaluation often relies on observations alone. We evaluate representative DGMs, flow matching (FM), DDPM, score-SDE, and VAE, on a known non-stationary Gaussian random field. This paper provides comprehensive metrics to assess recovery of the ground-truth mean and covariance structures, with oracle samples and a stationary control as references. All four models recover the mean surface, while their covariance recovery differs across model families: DDPM and score-SDE recover the covariance structure reasonably well, FM exhibits mildly attenuated non-stationarity and slight variance under-dispersion, and VAE has difficulty recovering the covariance structure. An experiment on ERA5 temperature anomalies further demonstrates how the framework can support the validation and development of DGMs for complex real-world spatio-temporal data.