Project 4

Project Location:

Clayton (VIC)

Desirable skills:

• Proficiency with Python using Linux-based systems
• Familiarity with principles of statistical machine learning
• Experience with modern ML libraries such as PyTorch

Supervisory project team:

Richard Scalzo, Gerald Pereira, Gary Delaney, David Howard, Rafael dos Santos de Oliveira, Edwin V. Bonilla, He Zhao, Daniel Steinberg, Dr. Dan Pagendam, Dr. Joel Dabrowski, Dr. Petra Kuhnert; Rui Wang and Ulrich Engelke

Contact person:

Project description:

AI4Design is a generalizable template for AI-assisted design of 3-D printable industrial components and processes, with the goal of accelerating development of new technologies in Australian industry for global competitiveness and transition to a green economy.  The two foundational AI-related problems at the heart of AI4Design are design generation — including unusual designs a human would not have come up with easily — and intelligent autonomous testing, calibration, and extension of advanced numerical simulations to better match real experiments.

The current AI4Design portfolio includes a variety of application projects around which to create cross-cutting capabilities.  While this work is ongoing, the PhD research topics will focus on fundamental questions involved in generalizing the AI4Design template for the long term to new areas beyond the current applied projects, particularly regarding coupling to physical models and advanced generative design.  Potential areas for this work include:

– Coupling of design representations to generative models such as diffusion models
– Learning of novel representations for design spaces that are compact and easy to explore without sacrificing complexity (for example using physical laws as constraints)
– Transfer learning for multi-physics-informed surrogate models to support new design capabilities for more complex problems

– Analysis & integration of novel high fidelity expt data such as tomographic imaging directly combined with multi-sensor measurements
– Development and application of novel XR systems to visualise complex multi-representational data from sensors, computational simulations and AI/ML models

Activities:  The students will work together using test cases aligned with applied AI4Design projects (including both direct and indirect representations, and with physics engines including CFD, DEM, FEM).  This cohort-based approach will ensure the students share knowledge across design problems useful to produce general tools.  Activities include:

• Literature review on related areas (generative deep learning, transfer learning, physics-informed learning)
• Implementation at a simplified level for a highly constrained or low-dimensional problem
• Development of a Bayesian experimental design platform, which will combine multi-fidelity simulation models to solve sequential experimentation problems involving multiple objectives, high throughput, and mixed, possibly high-dimensional parameter spaces.
• Testing on a case closer in complexity to an AI4Design application, or on an actual AI4Design application if suitable

Outcomes:  These will include new frameworks for AI-assisted design across a broad range of applications.  The principles would feature in high-impact publications while the details of software implementation would contribute to the AI4Design platform.