07 Apr Crowdsourcing Fusion: Prediction Scoring Leaderboard for General Fusion’s Data-Driven Challenge
for Submissions up to 2016 Apr 06, 11:59 PM EST
General Fusion’s crowdsourced challenge, titled “Data-Driven Prediction of Plasma Performance”, is open through April 13th and is hosted by InnoCentive. Details and registration information are available at https://www.innocentive.com/ar/challenge/9933692.
With the challenge closing next week, this will be the final weekly leaderboard posted.
In this challenge we are asking solvers to apply statistical techniques and computational tools to identify new patterns in hundreds of gigabytes of our data from plasma experiments. Plasma is the super-heated hydrogen gas that fuels General Fusion’s Magnetic Target Fusion process. Every five minutes or so, our plasma injectors create a compact toroid plasma that lasts for a few hundred microseconds – we call each of these “shots”. We’re eager to see if the solvers can identify patterns in the data that will allow us to further improve the quality and performance of our plasma (for example, making the plasmas last longer).
As an example of the kind of insights we’re looking for, consider the graph below:
This graph shows the magnetic field strength (kind of like magnetic energy) of two plasmas over time and paints a good picture of how long each plasma lasts. These are two consecutive shots on the same plasma injector, so the condition and configuration of the injector is mostly the same, but one shot lasts almost twice as long as the other. Why?
In some cases, we know exactly what was different between any two such shots, but in other cases we’re still missing a few pieces of the puzzle. We think that there might be some clues buried in hundreds of gigabites of data collected from hundreds of different sensors over thousands of plasma shots, and we’re looking to the crowd to help find these clues. In particular, 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.
As is described in the full details of the challenge, we’ve asked solvers to build a method to predict the performance of a plasma shot using only the data in the first 100 microseconds of the shot. We’ve provided this kind of data for a few hundred shots and are asking solvers to submit their predictions for those shots in order to score or rank their algorithms. Solvers can submit their algorithm’s predictions once a week for ranking in our weekly standings. This means solvers can keep their eye on the competition, refine their predictions, and resubmit to climb the next week’s standings.
This week’s leaderboard is given below. With the challenge closing next week, this will be the final posted leaderboard, though solvers have a little more time to refine their proposed solutions before next week’s submission deadline. Solvers are reminded to submit their final proposed solutions to InnoCentive via the challenge webpage.
Over the last few months, we’ve certainly seen improved scoring and plenty of competition, and we’d like to thank all of those who have participated.
* As a reminder, predictions are scored according to the following formula, with the objective being a minimal score: Where n is the number of plasma shots requiring predictions, p_i is the submission’s predicted performance for the i-th plasma shot, and, a_i is the actual experimental performance for that shot (known only to General Fusion).
** The root mean square (RMS) of the error for a solver’s most recent score is also provided to give additional context.
Solvers are reminded that they may ask questions or seek clarifications from InnoCentive through the challenge webpage.
Media inquiries about the challenge or General Fusion’s Open Innovation program should be directed to: