[Paper Review] Particle Filtering and Gaussian Mixtures -- On a Localized Mixture Coefficients Particle Filter (LMCPF) for global NWP

Anne Rojahn, Nora Schenk+2 more|arXiv (Cornell University)|Jun 15, 2022

Climate variability and models4 citations

TL;DR

This paper proposes the Localized Mixture Coefficients Particle Filter (LMCPF), a novel data assimilation method that integrates Gaussian mixture modeling into the Localized Adaptive Particle Filter (LAPF) to improve particle movement toward observations in global NWP. By modeling individual particle uncertainty with localized Gaussian mixtures, the LMCPF enables more accurate, non-Gaussian posterior updates, achieving analysis and forecast quality comparable to the operational LETKF in a 52 km global ICON model setup with 10⁶ variables.

ABSTRACT

In a global numerical weather prediction (NWP) modeling framework we study the implementation of Gaussian uncertainty of individual particles into the assimilation step of a localized adaptive particle filter (LAPF). We obtain a local representation of the prior distribution as a mixture of basis functions. In the assimilation step, the filter calculates the individual weight coefficients and new particle locations. It can be viewed as a combination of the LAPF and a localized version of a Gaussian mixture filter, i.e., a Localized Mixture Coefficients Particle Filter (LMCPF). Here, we investigate the feasibility of the LMCPF within a global operational framework and evaluate the relationship between prior and posterior distributions and observations. Our simulations are carried out in a standard pre-operational experimental set-up with the full global observing system, 52 km global resolution and $10^6$ model variables. Statistics of particle movement in the assimilation step are calculated. The mixture approach is able to deal with the discrepancy between prior distributions and observation location in a real-world framework and to pull the particles towards the observations in a much better way than the pure LAPF. This shows that using Gaussian uncertainty can be an important tool to improve the analysis and forecast quality in a particle filter framework.

Motivation & Objective

• To address the limitation of the LAPF in pulling particles toward observations when prior distributions are distant from observations.
• To improve the non-Gaussian analysis capability of particle filters in high-dimensional global NWP by incorporating model and forecast uncertainty per particle.
• To develop a stable, operational-grade particle filter framework that handles non-Gaussianity and high dimensionality in global atmospheric modeling.
• To evaluate the LMCPF’s performance against the operational LETKF in terms of analysis error, forecast skill, and ensemble spread stability.
• To investigate the impact of particle uncertainty modeling on the size and direction of particle moves in the assimilation cycle.

Proposed method

• The LMCPF extends the LAPF by modeling the prior distribution as a localized mixture of Gaussian basis functions, where each particle carries its own mean, covariance, and weight.
• It computes posterior mixture coefficients, means, and covariances using a Bayesian update that accounts for observation errors and localization.
• The filter uses adaptive resampling and rejuvenation techniques from LAPF to maintain particle diversity and prevent filter degeneracy.
• Observation localization is applied to the Gaussian mixture components to reduce sampling errors and improve scalability in high-dimensional systems.
• The method incorporates particle-specific uncertainty via a localized kernel-based approximation, inspired by LRKPKF and Liu et al. (2016).
• The analysis step computes particle weights based on the Mahalanobis distance between observations and projected ensemble means, adjusted by individual particle covariance.

Experimental results

Research questions

• RQ1 Can the integration of Gaussian mixture modeling into the LAPF framework improve particle movement toward observations in global NWP?
• RQ2 How does the inclusion of particle-specific uncertainty affect the stability and accuracy of the analysis in high-dimensional systems?
• RQ3 To what extent does the LMCPF achieve analysis and forecast skill comparable to the operational LETKF in a global 52 km resolution model?
• RQ4 How do the size and distribution of particle moves depend on the uncertainty parameters of individual particles?
• RQ5 What is the impact of localization and mixture coefficient adaptation on the ensemble spread and error statistics?

Key findings

• The LMCPF achieved stable, month-long global assimilation runs in a quasi-operational setup with 10⁶ model variables and 52 km resolution.
• The LMCPF reduced upper air first guess errors by 1–3% compared to LETKF below 850 hPa, demonstrating improved performance in the lower troposphere.
• Forecast RMSE for temperature and wind fields was comparable to LETKF, with differences within ±2–3% over a one-month period.
• The mean particle move size was analytically linked to the eigenvalue distribution of the metric tensor used in the weight computation, explaining observed histogram shapes.
• The ensemble spread of the LMCPF was stable after a spin-up phase and strongly dependent on tuning parameters, indicating sensitivity to configuration.
• In short-range case studies, the LMCPF outperformed LETKF in first guess RMSE, showing potential for improved initial condition quality.