I am using a solver to reduce the simulations of Miranda fusion reactors, as long as it has a lot of input data and needs some seconds to simulate (I use C++ and 8 threads)
My idea is that the algorithm generates some input datasets, obtain results and using the results throw new datasets in the better conditions.
I am trying a genetic algorithm to automatize simulations but needs a lot of unuseful simulations to work.
I think a more sophisticated method that obtains a more useful datasets
I added the result of one of the datasets.
Thank you Aparna. My simulator is written in C++ and those libs are in phyton, how I should make them work together?
best regards
I have seen that methods that are based on the Jacobians and inverse functions. The problem in the simulations are that there are a lot of local minimums to be obtained, and I do not know if those methods works well with local minimums.
In the following plot I obtained the result for y(x0) where x0=x1=x2... that are the different variables (or dimensions). There are results for 2 different configurations. The solver used by excel works well sometimes:
Ok, it seems that I need, I would like have more python solvers to add to the project. I should need one based on a mix of evolutionary algorithm with some using slopes or partial derivatives
Now I managed to improvethe confinement by using some equations to reduce the iteration number. I have found that tokamaks works well using an even number of coils, and by alternate current direction to avoid the plasma going to the upper (or lower) side.
But still it is very difficult to confine >500keV plasma and a lot of computation power is required, I should need to hire a computer cluster. It is possible using a c++ project or must be a python one?. I need using the boost library.
Can the root do it? (I think ellint functions could came with c++)
I improved the Aeff that is the averaged reactor section by using the "23 fellow system", which involved to generate vectors of parameters with a variation between them (it is a genetic algorithm as said at http://dprogrammer.org/ ) but when having more than 23 vectors the worst of them is erased, then the percentage of variation is reduced every time a new vector is obtained as a variation of the 23 fellows.
With that system it is reduced (and improved) the area from 2e-3 (m2) to 200e-4 in only 80 iterations, but unfortunately it is hard to reduce further. As it is said in the dprogrammer web, the algorithms that improves a lot at the beginning are worse in the long term.
Actually I am running the computer 12 hours every day simulating between 2000 and 5000 tokamaks every day
The decrease of the cross section of the reactor is a 10% lower than initial after 1400 tokamaks simulation:
https://www.researchgate.net/project/PULSOTRON-500k-SE-500KeV-THERMONUCLEAR-FUSION-REACTOR/update/5cfcf1eecfe4a7968da8bd95
After only 50 simulations it begins to give much better results than 5000 simulations using a random genetic algorithm.
The problem is that the better reactors takes 10-20 minutes per reactor and the bad ones can be tested at 10 reactors per second so the method should be carefully to select the good ones to avoid stuck in similar devices
At the end I had to use a genetic algorithm, but as long as I have to modify 5-6 variables I had to use the best between 1000 simulation results for next iterations to avoid lose a lot of possibilities. Using that I had to simulate >140000 reactors, after reactor number 20-30000 they did not reached ignition conditions. I am simulating about 10k-30k reactors every day.
Here is the result of one of them:
https://youtu.be/lzolQNfYFxw
best regards
- Badges
- Science topic