AI, ML and Friends is a weekly seminar series within the School of Computing on Artificial Intelligence, Machine Learning, and related topics. We are open to attendees and presenters external to the school. Please sign up to the mailing list to receive weekly announcements including zoom details, and email the seminar organiser to schedule a talk.
Upcoming Seminars #
21 November 2024, 11:00 #
Physics-Informed Neural Models for Grad-Shafranov Equation with Formal Verification #
Speaker: Fauzan Nazranda Rizqan
Abstract: Fusion energy offers promising potential as a nearly limitless, safe, and environmentally friendly power source. Central to achieving this is maintaining plasma equilibrium, where the balance between plasma pressure and confining magnetic fields must be maintained for stable operation. This equilibrium is described by magnetohydrodynamics (MHD). In axisymmetric tokamaks, this can be mathematically modeled using the Grad-Shafranov Equation (GSE) through the assumption of toroidal symmetry. In recent years, neural networks, particularly Physics-Informed Neural Networks (PINNs), have gained attention as a tool for modeling complex physical systems by embedding known physical laws into the learning process. However, current implementations of PINNs for solving the GSE have not been explored in settings in which one model is used for evaluating arbitrary fixed boundaries. Thus, we developed a new PINN that handles arbitrary fixed boundaries by including boundary points as network inputs. Additionally, we explored Fourier Neural Operators (FNOs) as an alternative for modeling the GSE, comparing its inference speed and computational efficiency with the PINNs. While both methods showed similar accuracy, PINNs provided much faster inference times, along with shorter training durations. Furthermore, we propose a novel approach by evaluating the amenability of the model-checking tool Marabou, to the task of verifying the adherence of trained PINNs to key physical properties. In some situations, we were able to observed a discrepancy between the model evaluations in PyTorch and according to Marabou. Nonetheless, we found that Marabou can still effectively verify key physical properties. It shows strong potential in validating constraints and identifying physical parameter space, especially for proving properties through counterexamples. Together, our findings demonstrate the potential of using physics-informed neural models for solving the GSE, with the addition of a feasibility test for formal verification using Marabou. This latter development is entirely novel for PINNs, and shows great promise. The use of PINNs for real time control application and surrogate modelling of complex physical processes in fusion plasmas (e.g. turbulence) is rapidly expanding. Model checking provides an opportunity to identify regimes of control system failure and unphysical surrogate model predictions.
Bio: Fauzan is a final-year undergraduate student pursuing a Bachelor of Advanced Computing (Honours) at the Australian National University, specialising in Machine Learning. He has numerous experiences in machine learning and software engineering through various projects, including a successful Data Visualisation project for the ANU Techlauncher and an Honours Thesis written under the supervision of Matthew Hole and Charles Gretton. This thesis focuses on an evaluation of neural models for a well-known equation in the field of plasma physics called the ‘Grad-Shafranov Equation,’ which is pivotal for understanding magnetohydrodynamic (MHD) equilibrium in fusion devices. Fauzan’s work also introduces a novel feasibility test of neural network verification tool for Physics-Informed Neural Networks (PINNs).
Where: Building 145, room 1.33
28 November 2024, 11:00 #
###
Speaker: Songtuan Lin
Abstract:
Where: Building 145, room 3.41