The deep latent space particle filter for real-time data assimilation with uncertainty quantification.

Area of Science:

• Computational Science
• Data Assimilation
• Machine Learning

Background:

• Data assimilation combines observations with simulations for accurate system estimation.
• Current methods are computationally expensive, hindering real-time application in complex systems.
• Accurate uncertainty quantification remains a significant challenge.

Purpose of the Study:

• Introduce a novel particle filter methodology, the Deep Latent Space Particle Filter (D-LSPF).
• Overcome computational limitations of traditional data assimilation techniques.
• Enable real-time data assimilation with uncertainty quantification for complex physical systems.

Main Methods:

• Developed the Deep Latent Space Particle Filter (D-LSPF) utilizing neural network-based surrogate models.
• Employed Wasserstein Autoencoders (AEs) with vision transformer layers for dimensionality reduction.
• Utilized transformers for parameterized latent space time stepping.

Main Results:

• D-LSPF demonstrated significant speed improvements, running orders of magnitude faster than high-fidelity particle filters.
• Achieved 3-5 times faster performance compared to alternative methods.
• Showcased up to an order of magnitude improvement in accuracy.
• Successfully applied to leak localization in multi-phase pipe flow and seabed identification.

Conclusions:

• The D-LSPF effectively addresses the computational cost of data assimilation.
• Enables real-time data assimilation with uncertainty quantification for tested complex systems.
• Represents a significant advancement in applying data assimilation to real-world problems.