Huh – no standing ovation yet… you didn’t mistake me for Tony Stark?  Or Ironman?  

Ok, you’re right, I don’t fight super villains or have a super cool suit.  But I think you’ll find I am more like Ironman than you realized. Let’s recall a scene from one of the Ironman movies, Tony Stark designs and creates a particle accelerator in his lab with his Artificial Intelligence assistant, Jarvis…. That scene describes my office!  (sort of) First let’s de-movify this, starting with particle accelerators.

Charged particle accelerators create beams of subatomic particles and accelerate them to (almost) the speed of light. These beams can then be shot at tumors to treat cancer, smashed together to reveal the laws of physics, used to create powerful light sources, and much more!

What do we do if we need another particle accelerator, maybe a more advanced Advanced Light Source?  First we have to design it and model the design via computer simulations. Particle accelerators are complex machines, with hundreds of intricate elements that have to be precisely aligned. Their design and modeling is equally involved. In some advanced accelerator concepts, just one component can require the largest supercomputers to be modeled accurately. Then we want to stack these together just right, and run this through the computer over and over again.  This is long and gets really expensive! How can we speed up expensive computational design of particle accelerators?

Accelerator scientists, including my group here at Berkeley lab, are using the Marvel superhero and inventor Tony Stark, also known as Ironman, as an inspiration to solve this problem. Flash back to the scene in the Ironman movie where he builds and runs a particle accelerator in his lab with the assistance of an artificial intelligence (AI) agent called Jarvis.  Like Ironman, many scientists are developing various AI tools to improve accelerator design. 

My Jarvis-like assistant  is made of neural networks.  I use neural networks to create models of advanced accelerator elements.  Neural networks are networks of simple mathematical functions that combine to learn elaborate mathematical functions.  I train my neural networks to learn from the data generated in computer simulations of advanced accelerator elements. 

I hope my neural networks help me catch details in the high-fidelity simulation data I might have missed and can then accurately predict what happens through arrangements I haven’t simulated yet.  And I hope my networks  do this fast!  This work is going on now, but already the neural networks can make accurate predictions, thousands of times faster than the simulations they were trained on! My neural networks are looking like promising candidates as Jarvis-like AI assistants for accelerator design. 

Maybe I can’t save you from evil villains.  But I and scientists like me and our ai assistants are designing particle accelerators to bring you more cancer treatments, more scientific discoveries, and more light; and that is pretty heroic.