Each experiment that TAE runs is incredibly complex. To start, physicists select which of the thousands of parameters on the machine they plan to test. Then they inject gas at the two ends of a cylindrical machine and ionize it to create plasmas. Physicists accelerate the two plasmas toward each other using powerful magnets that surround the walls of the machine. When the plasmas collide, they use magnetic and electric fields to try to keep the resulting plasma stable.
Each of these steps is, of course, dependent on the physicists’ selections, but the whole experiment is further complicated by the general instability of plasma. Even in a highly specialized machine like Norman, the plasma can tilt, wobble, or bounce into the walls of the cylinder. And in these cases, the physicists need to know what happened inside the machine in order to interpret the results of the experiment. As a result, they spend a good deal of time on an exercise they call “plasma debugging”—using sparse data from the machine’s sensors to try to infer the internal state of the plasma.
This is one of the processes where John and his team believe machine learning can make a big difference. Using TensorFlow, they’re building a plasma debugger that will more efficiently and more accurately provide physicists with information about the plasma state, including inferences about its density, shape, temperature, magnetic field, and more. To do this, the debugger takes the information from the machine’s sensors and maps it against a 3D simulated plasma model. It then introduces another layer of priors (things we know to be true) based on the laws of physics and the constraints of the experiment. Using all of this information, the debugger can produce a distribution of possible plasma states for TAE to review. In under five minutes, it allows physicists to explore different scenarios, make inferences about the experiment, and move on to their next trial. This keeps the pace of experimentation high and allows physicists to learn as much as possible from each test they run.