Machine Learning or Deep Learning for a Lab Experiment

[jamie.j1989]

Hi, I want to try out a bit of machine learning or deep learning with an optimisation problem in the lab. However, I'm confused at what the best option would even be or whether my optimisation problem is even applicable to either.

Firstly, the lab set up hasn't been built yet, I am computing the outcome. The model consists of an analytical solution to the magnetic field produced by current loops separated by some distance along the same axis. These current loops produce a quadrupole field where atoms for the proceeding experiment can be trapped at the zero-field region. However, first, the atoms must be transported down the common axis of the loops by ramping varying currents through these loops. Which amounts to moving the magnetic field-zero along the same path. There are various constraints along the way, such as the maximum currents in the loops the minimum field gradient at the zero-field and the time taken.

Does this sound like a problem that can be optimized with either machine learning or deep learning? I'd like to get into it if so as there are many experimental sequences that could be neatly optimised.

Thanks

 

[.Scott]

It's not clear to me what you are attempting to solve. Are you looking for the positions of the current loops? Or are those fixed and you are looking to control the current through those loops?

To the extent that I understand what you are doing, it sounds like you can model any specific current setup - and you want to discover the best sequence to follow to place the trapped particle at their destination.

Before the lab is finished, it would make sense to model this process; attempt varying strategies, modelling each; and them close in on the most optimal one.
That would not necessarily be considered "machine learning", but that term is so broad, it could be.

Once the lab is working, you can take actual results and build that into you model. Here, "machine learning" could be either the refinement of the model or the automated trial-and-error determination of an optimal transport sequence - or both.

 

[jamie.j1989]

Yes, current loops are static, varying the currents in the loops control the location of the magnetic field zero.

I can see how I should vary the currents to accomplish what I need, It would just be a nice project to see how I could set up some sort of machine learning process that could discover a more efficient route. And I'd like to implement it for future use with more complex parameters, such as trying to maximize the total atom number at the end of the sequence, minimize the heating of the atoms during the transport etc which would require actual feedback from the experiment. However, for now, I don't see why I wouldn't be able to feed in computational results to play around with and get a feel for what is needed.

Specifically what I now have in mind is having a predetermined path for the magnetic field zero as a function of time, which I will decide. This is then the 'goal' that whatever machine learning method I might be able to use should aim towards by optimally selecting the correct currents in the loops at some time value. I'm just unsure what method is best suited for this type of optimization?

 

[.Scott]

@jamie.j1989:
There is probably a way to use machine learning for this, but I would try a more direct approach.
Let's say you have 6 currents you are controlling. You should be able to adjust them to get the zero at a particular spot - somewhere along the path you will want it to follow later.

So let's say you find that this setting: 1.1, 2.2, 3.3, 4.4, 5.5, 6.6 gets you to your starting point.
I would next make fine adjustments to each of these current settings and watch how the zero point moves. In each case, a small delta (say 0.01) will cause a change in the position. So you will have (I for current) a dX/dI, dY/dI, and a dX/dI for each of the six loops.

Now, if you have a way of automatically measuring those changes in position, you could program the device to automatically collect this data. It could then use this to find the minimum current solutions for the entire path. That type of self-calibration is considered machine learning. If you cannot collect the zero position automatically, then you cannot take yourself out of that calibration loop.