Fuchs v1 Optimization

Banana Plot Gifs

Since last time, Chris told me I should try plotting the banana of points within the minimum value of the objective function. Below, I have a gif showing how the parameter space evolves in time (sorry for bad quality) for the best 5 percent.

I assumed maximum laser intensity and explored how the focal distance and target thickness would need to change for different values of $K$. Below, is another GIF with looking at just the best 1 percent.

Then, I proceeded to make the same plots but modifying the objective function to include a penalty term so that if the max proton energy is more than 15 percent from the cutoff energy, a large number is added to the objective function. Here, I have within 5 percent.

and here I have within 1 percent

A Second Attempt

I modified the objective function to look like the following

\begin{equation} f = \frac{\lvert E_\text{MAX} - E_C \rvert}{E_C} + 0.1 \eta / \text{EFF} + \text{PENALTY} \end{equation}

where EFF = $E_\text{TOT}/E_\text{L}$ is the proton conversion efficiency and PENALTY = 1000 if $E_\text{MAX}$ is more than 15 pct away from $E_C$. I made some python programs to compute the SVR, GPR, and NN predictions for 100000 random points (forcing intensity to be a maximum at 1e19 Wcm2) and storing them to a csv file. That way, I can use pandas to apply the objective function and sort very quickly to generate the “banana” plots. Below, I have an example of one of these plots

Here, I made the energy cutoff 0.5 MeV and the $\eta$ parameter equal to $0.01$ which has the interpretation of normalizing to a 1 pct conversion efficiency. In the plots, I highlighted the optimal point predicted by each model and the region within the best 5 pct of the optimal value predicted by each model. We can see that the SVR and GPR both underpredict the target thickness slightly from the optimal value when there is a lot of noise. However, the shape of the region is still close.

When there is no noise, the predictions look pretty spot on for SVR and GPR which is a good sign that the models are working correctly.

NIF/JLF Conference

I attended the NIF/JLF Users Meeting Conference from 01/30-02/01. There were a lof of talks: many talking about diagnostics at the NIF and about various experiments being performed there. I presented the same poster I presented back at LaserNet23 and the HED summer school. It was nice having the opportunity to attend the HED summer school, because some of them ended up attending the conference. I