Physics-informed Machine Learning Approaches for Numerical Modelling of Structures Under Extreme Conditions

Project description

The Australian government’s State of the Environment report highlights the increase in both the frequency and severity of extreme events due to climate change. Concurrently, populations and urbanization are rising in Australia and India. This combination necessitates a focus on infrastructure resilience against extreme loads. Large infrastructure projects, like the Cross River Rail in Australia and the Great Nicobar Development in India, require advanced numerical modelling at the design stage. However, accurately modelling structures under extreme loading scenarios, such as fires, earthquakes, and typhoons, is technically challenging and costly. This is due to the need for detailed 3D discretization to capture localized failure modes like lateral torsional buckling and macro-cracks. These detailed models are computationally expensive, as opposed to using cheaper beam-column finite elements for the structural frame and limited shell finite elements for floor slabs.

This project aims to develop a physics-informed machine learning approach for enriching inexpensive beam-column and shell finite elements with the ability to capture localised failure modes. Machine learning (ML) models will be trained against detailed finite element models to capture the complex relationships between material properties, stress state, and localised failure modes. Based on existing literature and prior experience, it is expected that tree and gradient-boosting algorithms will perform best for this task. They will be trained to correlate end-displacements with localised failure response. After that, the best-performing trained models will be coupled with simple beam-column and shell finite elements that provide a physics-based input for the models to use for assessment of the localised failure modes. In this way, the complex and difficult-to-predict failure modes will be captured by the ML model, while the overall adherence to equilibrium and laws of mechanics will be enforced by the global system of equations that is generated from a global model using the simple finite elements.

Outcomes

This project is divided into four primary work packages:
WP1: Evaluation of critical localised member-level failure modes using detailed finite element models.
WP2: Development of a series of machine-learning algorithms and training them on the detailed model from WP1.
WP3: Evaluation of the computational expense of the ML models from WP1, and coupling the most efficient few with the simpler beam-column and shell elements.
WP4: Validation, profiling, and optimisation of the finite elements enriched with physics-informed ML models.
The expected outcomes of the project are as follows:
1. Database of highly detailed finite element models assessing localised failure of individual structural members under extreme earthquake and fire loading.
2. A set of trained ML models for predicting localised failure in structural components.
3. The development of a hybrid Physics-ML finite element for computationally efficient analysis of failure of large structures under extreme loading conditions.

Essential capabilities

Knowledge of Structural Engineering, Mechanical Engineering (thermodynamics), programming, and good communication skills

Desireable capabilities

Experience with finite element modelling and background in Machine Learning

Expected qualifications (Course/Degrees etc.)

An eligible degree in Structural Engineering or Fire Safety Engineering or Mechanical Engineering