09 Dec Crowdsourcing Fusion: Data-Driven Prediction of Plasma Performance
Earlier this year we at General Fusion ran our first crowdsourced challenge. The hope was to find a clever new way to seal components of our fusion system in a manner that better survives extreme impacts, pressure and heat. Kirby Meacham, an MIT-trained mechanical engineer with his name on 35 US patents and about as many years of experience, claimed the $20,000 prize for his “Metallic Pressure-Balanced Anvil Seal” design.
What exactly is a crowdsourced challenge? Simply put, we describe a technical problem for which we need an innovative solution, put out a cash prize for the best solution(s), and send it out to the entire world. The idea is to tap into the global wealth of knowledge by looking outside of our organization for ideas from a wide range of industries. After all, while we have many talented, smart people here at General Fusion, we know that there are thousands of others around the world. We believe that crowdsourcing is one of many tools that will help General Fusion on our mission of creating abundant, clean, safe and affordable energy.
When we announced the award, we promised more challenges and hinted that one would look to the crowd for new insights into General Fusion’s experimental plasma physics data. True to our promise, today we are excited to announce our second crowdsourced challenge, “Data-Driven Prediction of Plasma Performance”.
For this challenge, we are opening up the experimental data collected from a series of tests using PROSPECTOR, General Fusion’s best-performing plasma injector which has created the longest-lived compact-toroid plasmas of their type. We’re asking participants to apply statistical techniques and/or computational tools to identify new patterns in our data.
When operating, every five to ten minutes PROSPECTOR created a single compact toroid plasma which lasts for anywhere between a few hundred microseconds to over a millisecond – we call each of these “shots”. Our shot data includes signals from nearly one hundred probes measuring things like magnetic field strength, plasma density and the spectral composition of plasma light. There are also configuration settings for each shot, and calculated single point, or scalar, metrics.
Most importantly, for each shot we provide a single calculated value that quantifies the overall quality of each shot. We’ve learned that even when we apply all of the conditioning and operational techniques we’ve developed over the years, the plasma performance varies from shot to shot. We can predict and understand some of this variability, but not entirely.
If this is starting to sound more like an advanced physics lesson than a description for a data analysis challenge, don’t worry! We’re not looking for participants who are plasma physicists and aren’t looking for a physics analysis of the data. Rather, we’re looking for participants who can find statistical patterns in large amounts of difficult to digest data, just like is done with data of all kinds, all over the Internet, every day.
We’re asking participants to build a method to predict the performance of plasma using only the data from the first 100 microseconds of a plasma shot (again, PROSPECTOR shots can last over 1000 microseconds). To facilitate this, we’re providing two data sets: one with all shot data, the other with only the first 100 microseconds of data. Participants will use the first set to develop their techniques and submit their predictions of the performance metric for the second set. Of course, we know the actual results so we can evaluate submissions based on the accuracy of their predictions and also on the description of their methods. The best prediction method will earn the top prize of $20,000.
Understanding what parameters predict plasma performance can help guide the design and operation of future plasma injectors, and we anticipate that this challenge will give us some great insights in this area.
Participants can submit predictions once a week and to help fire up the competition we’ll post weekly standings. This way everybody can keep their eye on the competition, refine their predictions, and resubmit to climb the next week’s standings.
As with the first challenge, this challenge will be hosted by InnoCentive, who have conducted similar successful challenges with industry leading organizations such as NASA and Procter & Gamble. All details of this challenge and eligibility requirements can be found on InnoCentive’s website.
We’ve designed this challenge to give a broad range of people across the planet a chance to work with us to solve exciting problems with world-changing potential impacts, and we’re excited to see what comes next.