[Paper Review] Adversarial autoencoders and adversarial LSTM for improved forecasts of urban air pollution simulations

César Quilodrán Casas, Rossella Arcucci+3 more|arXiv (Cornell University)|Apr 13, 2021

Wind and Air Flow StudiesEnvironmental Science18 references19 citations

TL;DR

This paper proposes a two-stage deep learning framework combining a PCA-based adversarial autoencoder (PC-AAE) and an adversarially trained LSTM to improve long-term forecasts of urban air pollution from CFD simulations. By enforcing latent space consistency via adversarial training, the method reduces forecast divergence from physical models, achieving 4 orders of magnitude faster predictions than CFD with superior accuracy and physical realism over 80 time-steps in South London simulations.

ABSTRACT

This paper presents an approach to improve the forecast of computational fluid dynamics (CFD) simulations of urban air pollution using deep learning, and most specifically adversarial training. This adversarial approach aims to reduce the divergence of the forecasts from the underlying physical model. Our two-step method integrates a Principal Components Analysis (PCA) based adversarial autoencoder (PC-AAE) with adversarial Long short-term memory (LSTM) networks. Once the reduced-order model (ROM) of the CFD solution is obtained via PCA, an adversarial autoencoder is used on the principal components time series. Subsequentially, a Long Short-Term Memory network (LSTM) is adversarially trained on the latent space produced by the PC-AAE to make forecasts. Once trained, the adversarially trained LSTM outperforms a LSTM trained in a classical way. The study area is in South London, including three-dimensional velocity vectors in a busy traffic junction.

Motivation & Objective

• To address the rapid divergence of standard RNN-based surrogate models from physical CFD solutions when forecasting over long horizons.
• To enhance data compression and forecast stability in reduced-order models (ROMs) of urban air pollution simulations.
• To leverage adversarial training to align the latent space of the model with a prior distribution, improving generalization and physical plausibility.
• To develop a data-agnostic, scalable framework applicable to diverse CFD models with sufficient training data.

Proposed method

• A PCA-based dimensionality reduction is applied to 3D velocity fields from CFD simulations, retaining the first τ principal components (PCs) to form a reduced-order model (ROM).
• A PCA-based adversarial autoencoder (PC-AAE) is trained on the PC time series, using a generator (encoder-decoder) and a discriminator to match the aggregated posterior of the latent space to a standard Gaussian prior.
• The reparameterization trick is used to enable backpropagation through the stochastic latent variables, ensuring differentiable training of the autoencoder.
• An adversarially trained LSTM (ALSTM) is then trained on the latent space produced by the PC-AAE, with the discriminator enforcing consistency between generated and real latent sequences.
• The ALSTM generates forecasts by autoregressively predicting future latent states, which are then decoded back into full-space velocity fields via the PC-AAE decoder.
• The framework is evaluated on a 3D urban air pollution simulation in South London with 100,040 nodes and 1,000 time-steps, using a 720m×676m×250m domain with realistic boundary conditions.

Experimental results

Research questions

• RQ1 Can adversarial training of an autoencoder improve the compression and reconstruction fidelity of principal components from CFD simulations?
• RQ2 Does adversarial training of an LSTM on the latent space of a PC-AAE reduce forecast divergence from the underlying physical model over long horizons?
• RQ3 How does the proposed ALSTM framework compare to classical LSTM and simple PC truncation in terms of forecast accuracy and physical realism?
• RQ4 Can the framework maintain performance across varying latent space dimensions (τ) without significant degradation?
• RQ5 To what extent does adversarial training enhance the generalization and robustness of surrogate models for urban air pollution forecasting?

Key findings

• The PC-AAE outperformed simple truncation of principal components in reconstruction error, with consistent performance across different latent dimensions (τ = 4, 8, 16, 32).
• The adversarially trained LSTM (ALSTM) reduced forecast error by 40% compared to a classical LSTM after 80 time-steps, with significantly lower mean absolute error and standard deviation in the ensemble forecast.
• After 80 time-steps, the classical LSTM forecast diverged markedly from the ground truth, while the ALSTM preserved key physical features such as vortices and eddies, as confirmed by visual inspection in Figure 3.
• The ALSTM achieved a 4 orders of magnitude speedup over full CFD simulations, enabling near real-time forecasting.
• The framework demonstrated robustness to latent space dimensionality, with optimal performance observed at τ = 8, achieving a 5-order-of-magnitude compression from full space to latent space.
• The discriminator effectively constrained the latent space, preventing out-of-distribution samples and improving long-term forecast stability.