There is a video of one of these experiments in the linked blogpost (figure 2), where we show imagery taken from a camera looking into the vessel. Next to it is a reconstruction of the shot, visualised on a cross-section of the torus.
The simulation not perfectly matching reality is indeed a major challenge, and the nature paper is mainly about showing how you can overcome that challenge. As others have said, plasma is a pain to model, and you cannot collect a ton of data on these machines. But it seems technology is coming at a point where we know how to work within these limits and still get it to work.
Do you think this line of research will eventually feed back into improved designs of electromagnets, chambers and so on? My limited understanding of the field is that the optimal reactor design is a moving target that's changed a lot over the years due to engineering improvements in manufacturing, material sciences and so on.
I'm imagining optimizing the plasma configuration, reactor shape, coil/sensor placement and controller all in one giant computational optimization. All the algorithmic components and models humanity needs to attempt this seem to be there.
Stellarators are particularly fun. First you optimize a last closed magnetic flux surface (a 3D sheet) with relatively simple plasma models, then you optimize coils (you need to weight things like magnetic field error, gaps big enough for heating and diagnostics, and coil complexity a la minimum radius), then you perform more accurate simulations to start the overall optimization loop again.
Episode 22 is an intro to fusion power research and tokamaks, but could be skipped if you are aware of the big concepts. Episode 157 is an interview of a director at ITER: the largest fusion power project in terms of scale, ambition, and funding. ITER is expected to demonstrate net energy gain fusion within the next ten years. Episode 157 gives a good sense of where the cutting edge of fusion power research exists currently and its potential place in society. Episode 304 interviews the authors of a recent popular science book on tokamaks. This episode covers the history of fusion power research, and the economics involved in future power plants. Episode 312 is a set of interviews at W7-X: the current largest stellarator. This episode take a deeper dive into what fusion research is like and how threads are being tugged on.
This same podcast also has interesting episodes on plasma physics and superconductors. The superconductor episode is especially worth listening to more than once.
Do you think this approach of controlling plasmas also translates to these types of reactors?
Reading your article I didn't understand a lot of things and wondered if there was some kind of fusion-reactor-physics word-list where the most important keywords are explained a little. (without reading a whole book on plasma physics :-))
Also the timescales are fascinating. How long is 200ms in a tokamak? How long does this reactor need to ignite or shut down the plasma?
So this is a good tool for the computationalists but we still need the experimentalists. This is science.
> So this is a good tool for the computationalists but we still need the experimentalists. This is science.
I would say that this is a good example of collaboration between the groups.
Not far off and you want perfect results? Go with actual data. Very far off where it's not even usable? You'd need to go back to the training environment and redo the reward function.
Generally speaking (and I'm not sure if this is the case from reading the link but I may have missed it) reinforcement learning for optimal control is done to fine tune while traditional control methods are used for coarse adjustments. Since this is deepmind they probably want to use RL for everything