Generalised Latent Assimilation in Heterogeneous Reduced Spaces with Machine Learning Surrogate Models

Sibo Cheng, Jianhua Chen, Charitos Anastasiou, Panagiota Angeli, Omar K. Matar, Yi-Ke Guo, Christopher C. Pain, Rossella Arcucci

Introduction

High-fidelity simulations, such as Computational Fluid Dynamics (CFD), are indispensable for modern engineering but are notoriously slow, often requiring hours or days to produce results. Machine learning (ML) surrogate models—fast approximations that operate in compressed "latent spaces"—offer a potential solution. However, these surrogates tend to drift from physical reality over time.

Data Assimilation (DA) is the standard technique used to correct model drift by integrating real-world sensor observations. The challenge in "Latent Assimilation" is that the model's compressed state and the sensor's observations often reside in different mathematical or physical spaces. Traditional methods struggle with the "jagged," non-differentiable mappings typical of deep neural networks, making real-time optimization nearly impossible.

The GLA Framework

To bridge this gap, this research introduces Generalised Latent Assimilation (GLA), a framework designed to correct fast-running ML surrogates with heterogeneous real-time data. The pipeline operates in three distinct stages:

  1. Dimension Reduction (POD AE): The system utilizes a hybrid of Proper Orthogonal Decomposition (POD) and Autoencoders (AE) to efficiently compress irregular mesh data from fluid dynamics into a latent space.
  2. Dynamics Prediction (LSTM): A Long Short-Term Memory (LSTM) network predicts the temporal evolution of the system within this reduced space.
  3. Local Polynomial Regression (PR): This is the core innovation of GLA. To map the simulation state to heterogeneous observations, the authors use local PR. By sampling the immediate neighborhood of a predicted state using Latin Hypercube Sampling (LHS) and fitting a smooth polynomial, they create a differentiable "surrogate of a surrogate" that enables rapid variational data assimilation.
Schematic of the Generalised Latent Assimilation (GLA) framework, illustrating the mapping between physical and reduced spaces and the role of the local surrogate function.

Experimental Results

The researchers evaluated GLA using a high-dimensional CFD simulation of a two-phase (oil-water) liquid flow. The model was trained on 1,000 snapshots of simulation data and tested against synthetic observations generated via non-linear functions to mimic complex real-world sensor behavior.

Key Performance Metrics

  • Massive Speedup: The ML surrogate model operated nearly 1,000 times faster than the original CFD simulation.
  • Drift Correction: Without GLA, LSTM predictions diverged from ground truth within 2 seconds. With GLA, the model successfully incorporated observations to maintain accuracy for the full 10-second test duration.
  • Handling Complexity: The system successfully processed non-linear observation operators (such as reciprocal functions) that standard Latent Assimilation methods typically cannot handle.

Discussion and Future Work

The use of local polynomial regression effectively "linearizes" the complex, non-linear mappings of deep learning encoders and decoders, making the optimization landscape for DA significantly more manageable. Furthermore, the study found that "joint models"—those predicting multiple physical variables like concentration and velocity simultaneously—were more stable, suggesting that preserving physical correlations in latent spaces is crucial.

While GLA offers a significant performance leap, it remains sensitive to hyperparameters such as the sampling range and polynomial degree. Future work will explore how uncertainty propagates from physical space into the latent space, potentially utilizing ensemble-type DA approaches to better manage noisy real-world sensor data.

Summary updated

BibTeX
@article{cheng2023generalised,
  author = {Sibo Cheng and Jianhua Chen and Charitos Anastasiou and Panagiota Angeli and Omar K. Matar and Yi-Ke Guo and Christopher C. Pain and Rossella Arcucci},
  title = {{Generalised Latent Assimilation in Heterogeneous Reduced Spaces with Machine Learning Surrogate Models}},
  journal = {J. Sci. Comput.},
  year = {2023},
  url = {https://openreview.net/forum?id=m6QMUt23v0D},
}
Figure