Comment on egusphere-2022-1167

Caili Zhang, Sibo Cheng, Matthew Kasoar, Rossella ArcucciOpenalex Unknown 2022 · DOI · OpenAlex

Abstract

The occurrence of forest fires can impact vegetation in the ecosystem, property, and human health, but also indirectly affect the climate. JULES-INFERNO is a global land surface model, which simulates vegetation, soils, and fire occurrence driven by environmental factors. However, this model incurs substantial computational costs due to the high data dimensionality and the complexity of differential equations. Deep learning-based digital twins have an advantage in handling large amounts of data. They can reduce the computational cost of subsequent predictive models by extracting data features through Reduced Order Modelling (ROM) and then compressing the data to a low-dimensional latent space. This study proposes a JULES-INFERNO-based digital twin fire model using ROM techniques and deep learning prediction networks to improve the efficiency of global wildfire predictions. The iterative prediction implemented in the proposed model can use current-year data to predict fires in subsequent years. To avoid the accumulation of errors from the iterative prediction, Latent data Assimilation (LA) is applied to the prediction process. LA manages to efficiently adjust the prediction results to ensure the stability and sustainability of the prediction. Numerical results show that the proposed model can effectively encode the original data and achieve accurate surrogate predictions. Furthermore, the application of LA can also effectively adjust the bias of the prediction results. The proposed digital twin also runs 500 times faster for online predictions than the original JULES-INFERNO model without requiring High-Performance Computing (HPC) clusters. The implementation code of this study and the developed models are available at https://github.com/DL-WG/Digital-twin-LA-global-wildfire.

Introduction

Predicting wildfires on a global scale is a massive computational challenge. Currently, scientists rely on complex tools like JULES-INFERNO, a Dynamic Global Vegetation Model (DGVM) that simulates how vegetation, soil, and climate interact to spark and spread fires. While accurate, these models are computationally "heavy." A 100-year climate scenario can take 17 hours to run on a high-performance computing (HPC) cluster using dozens of threads. This bottleneck makes it difficult for researchers to run multiple "what-if" scenarios quickly. Furthermore, when these models are used to predict the future iteratively (using this year's prediction to forecast next year), small errors tend to snowball, leading to "drift" and increasingly inaccurate long-term results.

Methodology: The Digital Twin Framework

To address these challenges, the authors propose a three-stage framework to build an efficient "surrogate" model:

1. Reduced Order Modelling (ROM)

To handle the high dimensionality of global climate data, the researchers used Convolutional Autoencoders (CAE) and Principal Component Analysis (PCA). These techniques compress massive 2D global maps into a tiny "latent space"—a vector of just 100 numbers—that captures the essential features of fire and climate patterns. The study found that CAEs significantly outperformed PCA because they are designed to understand spatial relationships on a map.

2. Deep Learning Prediction

They trained a Long Short-Term Memory (LSTM) network to predict how that 100-number latent vector changes over time. Instead of solving complex differential equations, the LSTM learns the temporal patterns of wildfire occurrences from historical simulation data.

3. Latent Data Assimilation (LA)

This is the key to long-term stability. Every five years, the model performs Data Assimilation within the compressed latent space. It compares its prediction against "observations" (encoded data from the original physical model) and nudges the prediction back on track. Doing this in the latent space is computationally inexpensive compared to traditional methods.

Key Results and Performance

The study utilized data generated by the JULES-INFERNO model, covering monthly snapshots from 1961 to 1990. The results highlight the efficiency and accuracy of the digital twin approach:

  • Massive Speedup: The digital twin performed online predictions 500 times faster than the original JULES-INFERNO model. It predicted 30 years of global fire data in just 35 seconds on a standard laptop, compared to roughly 5 hours for the physical model on an HPC node.
  • High Accuracy: The CAE with a 100-dimensional latent space achieved a remarkably low reconstruction error (MSE of 0.00016).
  • Error Correction: Without Latent Data Assimilation (LA), iterative predictions showed a clear drift from the ground truth. Applying LA every five years successfully stabilized the model and significantly reduced the average Mean Square Error (MSE) across all test scenarios.

Conclusion and Future Work

By reducing the requirement from an HPC cluster to a standard laptop, this approach makes sophisticated wildfire forecasting accessible to researchers and agencies without massive infrastructure. While the current model was trained on a specific 30-year window and used simulation data as a proxy for observations, future work will focus on integrating real-time satellite observations (like MODIS or VIIRS) and addressing potential systematic biases inherited from the base physical model.

The implementation code and developed models are available at: https://github.com/DL-WG/Digital-twin-LA-global-wildfire.

Summary updated

Abstract

The occurrence of forest fires can impact vegetation in the ecosystem, property, and human health, but also indirectly affect the climate. JULES-INFERNO is a global land surface model, which simulates vegetation, soils, and fire occurrence driven by environmental factors. However, this model incurs substantial computational costs due to the high data dimensionality and the complexity of differential equations. Deep learning-based digital twins have an advantage in handling large amounts of data. They can reduce the computational cost of subsequent predictive models by extracting data features through Reduced Order Modelling (ROM) and then compressing the data to a low-dimensional latent space. This study proposes a JULES-INFERNO-based digital twin fire model using ROM techniques and deep learning prediction networks to improve the efficiency of global wildfire predictions. The iterative prediction implemented in the proposed model can use current-year data to predict fires in subsequent years. To avoid the accumulation of errors from the iterative prediction, Latent data Assimilation (LA) is applied to the prediction process. LA manages to efficiently adjust the prediction results to ensure the stability and sustainability of the prediction. Numerical results show that the proposed model can effectively encode the original data and achieve accurate surrogate predictions. Furthermore, the application of LA can also effectively adjust the bias of the prediction results. The proposed digital twin also runs 500 times faster for online predictions than the original JULES-INFERNO model without requiring High-Performance Computing (HPC) clusters. The implementation code of this study and the developed models are available at https://github.com/DL-WG/Digital-twin-LA-global-wildfire.
BibTeX
@inproceedings{zhang2022comment,
  author = {Caili Zhang and Sibo Cheng and Matthew Kasoar and Rossella Arcucci},
  title = {{Comment on egusphere-2022-1167}},
  booktitle = {Openalex Unknown},
  year = {2022},
  doi = {10.5194/egusphere-2022-1167-rc2},
}
Figure