FindAPhD Weekly PhD Newsletter | JOIN NOW FindAPhD Weekly PhD Newsletter | JOIN NOW

Development of machine learning frameworks for predicting bushfire dynamics

   School of Mathematics

This project is no longer listed on and may not be available.

Click here to search for PhD studentship opportunities
  Prof Santiago Badia, Prof Julio Soria  No more applications being accepted  Funded PhD Project (Students Worldwide)

About the Project

Project description

Infrared satellite monitoring of bushfires provides incomplete and little information about their transport properties and dynamics, which complicates fire propagation predictions. This project will develop and apply a domain-specific scientific machine learning (SciML) framework in which a physics-informed machine learning will assimilate infrared satellite data to provide augmented information of the bushfire transport properties (e.g. wind velocity, convective energy transport) and their dynamics. The real-time generation of this augmented information is vital to provide real-time, high-fidelity predictions of bushfire propagation, which is fundamental to all aspects of the disaster management cycle of bushfires, including prevention, planning, response and recovery.

Experts involved in any of these stages rely on mathematical models of fire propagation. These models have been developed based on a small sample of controlled outdoor experimental studies for a specific fuel type and field observations. These models have limited applicability due to this small sample size and the lack of high-quality experimental quantitative data acquired under controlled experimental conditions for a given fuel type. Furthermore, they are not universal and do not apply to other fuel types with distinct fuel structure.

This project will employ a more general approach that integrates data assimilation of real-time satellite data with machine learning constrained by the governing conservation equations to deal with variations of fuel type, topography, etc. by leveraging the results from the Advanced Leadership Computing Grant (ALCG) awarded to Professor Soria to simulate the complex interaction between fuel sources with different properties and the atmospheric turbulent boundary layer during bushfires by using high-fidelity direct numerical simulation (DNS) of distributed high energy heat sources on a high Reynolds number turbulent thermal boundary layer (TTBL). This data will serve as the “ground truth” to train the SciML framework and evaluate its predictive abilities, as well as the uncertainty associated with the predictions.

The proposed project falls within MDFI’s “AI and data science in sustainable development” and the focus and impact challenges of the ages: “Climate change” and its consequences by, in the first place enabling dynamic monitoring, modelling and prediction of bushfires, and secondly with the long term objective of the development via SciML of efficient and realistic predictive dynamic models that will allow the design of sustainable farming and forestry plantation resilient to bushfires, such as appropriately designed fire-breaks based on scientific knowledge to reduce fire propagation and mitigate damage to fauna and loss of human lives.

This project is highly interdisciplinary bringing together the fields of turbulent thermo-fluids, direct numerical simulation of turbulent shear flows, scientific high performance computing (S-HPC), data assimilation and SciML that will investigate feed-forward dense neural networks, convolutional neural networks (CNN), generative adversarial networks (GAN), neural networks coupled with genetic algorithms, etc., as well as novel techniques not yet explored, which are subject to current research of the Professors Soria, Badia and their colleagues. The most successful technique will be identified via extensive experimentation using HPC platforms by leveraging a computational framework written in Julia which is general enough to accommodate all these techniques in a unified framework.

Given the relevance and importance of bushfires in Australia and other countries with similar vegetation and climate (for example, the Mediterranean countries, California), the outcomes of this project will be of interest to many stakeholders within this space, including the Victorian and other State governments, as well as the Federal government and beyond. Therefore, this project has a strong prospect for upscaling and securing direct funding from the Victorian Department of Environment, Land, Water and Planning, in addition to funding through ARC Linkage grants with the Victorian Department of Environment, Land, Water and Planning as the Partner Organisation.

PhD student role description

The role of the PhD student and their contribution to the project is to develop and validate the SciML framework to enable the high-fidelity predictions of bushfires based on experimental temperature field measurements from satellite imaging.

The student will contribute to the analysis and development of suitable machine learning-inspired discretisation of forwarding and inverse problems governed by PDEs. It involves the design of architectures fulfilling smoothness requirements, e.g. to end up with well-posed minimisation problems (differentiable loss functions with respect to neural network parameters. We will also explore suitable PDE residual terms in the cost functions. Furthermore, we will identify synergies between traditional grid-based and PINN-based schemes, e.g. to accelerate the training of the latter.

The data (i.e. the 3D temporally evolving temperature field) to test and validate the SciML framework will come from the simulation of an analog of a bushfire provided by the DNS of a TTBL with distributed high energy heat sources used to model the fuel sources, which is available and will serve as the “ground truth”. The development and validation of the SciML framework, which includes determining the optimal learning technique, such as feed-forward dense neural networks, convolutional neural networks (CNN), generative adversarial networks (GAN), neural networks coupled with genetic algorithms, as well as more novel approaches based on reduced order models deduced from the DNS of the TTBL, will be identified via extensive experimentation using HPC platforms, leveraging a computational framework written in Julia that is general enough to accommodate all learning techniques in a unified framework.

The PhD student will be part of a team which includes other PhD students from Professor Soria and Badia’s groups that are working on different aspects related to this project including the DNS of TTBL, using SciML to enhance experimental measurements of TTBL and to yield new knowledge of the dominant structures and their role in turbulent shear flows, the mathematical analysis of learning algorithms such as CNN, GAN, etc.

Required skills and experience

  • The candidate must have outstanding knowledge in Mathematics and/or the theory of Fluid Mechanics
  • Experience in programming (e.g. in Julia, Python, C/C++ or F90) is essential
  • Knowledge of MPI and/or MP is desirable
  • Experience in numerical simulations and/or machine learning is necessary

Expected start date: February 2023


Areas of research:
• Application of scientific machine learning (SciML) and AI for high-fidelity predictions of complex physical dynamics (i.e. bushfires) based on satellite imaging.
• Direct numerical simulation (DNS) of turbulent thermal boundary layer (TTBL) flow to enable the development and application of predictive SciML techniques based on experimental satellite imaging.
Search Suggestions
Search suggestions

Based on your current searches we recommend the following search filters.

PhD saved successfully
View saved PhDs