A data-driven approach to model the physics of superheated gas hitting a wall
Who is this presentation for?
- Anyone interested in applying machine learning to model natural and engineering phenomena
A big chunk of the world’s problems could be solved if we had access to nearly unlimited clean sustainable energy. Concerns of waste management and lack of clean water would be a thing of the past, and as energy prices drop, the global economy would be able to accelerate to unprecedented levels. To those working towards that dream, fusion energy is the final frontier, the holy grail of the energy race.
One of the big challenges impeding progress toward that goal is that we just don’t quite understand how to effectively ensure that the gas heated at tens of millions of degrees Celsius can be contained and prevented from tearing the reactor wall down. This puts physicists researching nuclear fusion in a conundrum. Without understanding how these superheated gases (plasma) interact with the wall, physicists can’t effectively attain fusion, and without running experiments that at the very least mimic fusion conditions, they’ll never be able to progress their understanding, either.
Over successive decades, physicists have designed simulation algorithms on top of physics models to help solve this conundrum. The idea is that these models will help physicists gain some insight into the processes that are undertaken when the plasma hits the reactor walls. But the complexity of the models has long been a hindrance in allowing physicists to make much headway. Employing these algorithms on the fastest supercomputers available would still take days, if not weeks, to converge and arrive at a solution.
In the field of fusion research, data is in abundance, and researchers have been running experiments and simulations for decades, accumulating petabytes of data. Based on this abundance, Vignesh Gopakumar uses a deep learning approach to infer physical functions inherent in the data. By switching from a physics-based model to a data-driven model that studies the evolution of plasma, he’s able to obtain a computational gain of five orders of magnitude.
Vignesh proposes a novel fully convolutional network (FCN) approach that would perform image mapping of the temporal evolution of the plasma across a two-dimensional cross-section of the reactor. The model is capable of predicting the future states of a physical variable such as temperature and density across the reactor. He engages in a dimensionality reduction approach to effectively capture the effective physics relations underlying in the images that characterize the distribution of these variables. The model maps with a certain degree of accuracy highly complex nonlinear representations while adapting to incorporate different physics models into it.
So far, he has been experimenting with such data-based models only within the simulation space. There’s substantial potential to build such models upon experimental data that has accumulated over 35+ years. He’s currently creating algorithms that can model the plasma behavior from such datasets and believes that maybe one day there might be AI-controlled fusion reactors.
- A basic understanding of the working of neural networks and its application as a universal function approximator
- Working knowledge of performing regression analysis using neural networks
What you'll learn
- Learn how to inferring physics principles from data
- Understand approaches to performing machine learning modeling of natural and engineering in data-bounteous scenarios
United Kingdom Atomic Energy Authority
Vignesh Gopakumar is a machine learning engineer specializing in fusion research with the United Kingdom Atomic Energy Authority. He spends his time building machine learning algorithms to model physics systems that help gain more understanding of the underlying phenomenons. He designs algorithms that help discover anomalies as well as predict malfunction of engineering systems. He’s working on building a model that can be augmented in real time when exposed to different physics principles.
For conference registration information and customer service
For more information on community discounts and trade opportunities with O’Reilly conferences
For information on exhibiting or sponsoring a conference
For media/analyst press inquires