Generative Model Proposal based Particle Filtering for Data Assimilation
2026-07-01 • Machine Learning
Machine Learning
AI summaryⓘ
The authors study a method for tracking changing systems using data and observations over time, known as data assimilation. Traditional methods often assume simple conditions that may not fit real-world cases, while existing advanced methods can fail, especially with complex or large-scale data. They propose a new approach called Flow Proposal Particle Filters (FPPF) that improves how particles are guided based on observations, leading to more accurate tracking. Their method keeps important Bayesian updating steps and works better in difficult scenarios. Tests show their approach performs well compared to other techniques in complex situations.
data assimilationfilteringparticle filtersBayesian updateconditional generative modelsoptimal proposal distributionweight variancelocalizationnon-linear systemshigh-dimensional data
Authors
Chandni Nagda, Mayank Shrivastavam Gudrun Thorkelsdottir, Gan Zhang, Morteza Mardani, Arindam Banerjee
Abstract
Data assimilation models state dynamics conditioned on sequential observations, and has wide-ranging scientific applications. In the filtering setting, the goal is to model the posterior over the current state given all observations so far. Classical solutions typically make simplifying distributional or functional assumptions, e.g., linear-Gaussian systems, which can be inaccurate in many scenarios. In principle, particle filters (PFs) remove these assumptions, yet often collapse in high dimensions. Recent generative approaches learn conditional state transitions, but without principled Bayesian updates they do not recover the correct filtering posterior and can accumulate error over long horizons. In this work, we introduce Flow Proposal Particle Filters (FPPF), which learn a conditional generative model based proposal approximating the variance-minimizing optimal proposal for particle propagation. Conditioning on observations steers particles toward high-likelihood regions before weighting, reducing weight variance and delaying degeneracy. Since our proposal admits tractable likelihood evaluation, FPPF computes accurate importance weights and retains a Bayesian update step. We further extend FPPF to high-dimensional problems through localization strategies, adressing another standard PF failure mode. Extensive experiments on a variety of dynamical systems show that FPPF outperforms statistical baselines and other generative methods in non-linear, non-Gaussian, and high-dimensional regimes.