This post covers work from 09/30-10/18. Now that I’ve done most of the work on the titan simulations, I decided to get back into the synthetic data paper which I’ve made good progress on. Additionally, I’ve been helping out with Nathaniel’s experimental paper: making some edits and suggestions.

Synthetic Data

I produced the main result which is shown below

I constructed a new objective function

\begin{equation} f(KE_c, \eta_p) = \frac{KE_c - KE_{c,goal}}{1 MeV} \beta - 100 \eta_p (1-\beta) \end{equation}

where $KE_c$ is the maximum proton energy which we call the “energy cutoff” and $\eta_p$ is the laser to proton conversion efficiency (total proton energy / laser energy from gaussian beam formula). The contour plots above show the regions where the objective function is minimized for a specific cutoff of 1 MeV and varying levels of $\beta$. The top row is the exact fuchs function with 0 noise. The other three rows are from models trained on 30 percent noise and we are seeing their predictions for a subset of the data that has a fixed max intensity and min contrast. We can see that the Neural Network most closely approximates the true FUCHS function, but it also looks a little blurry. I think the perspective here is that the Neural Network is more effective as a blackbox model. The polynomial regression is limited in what it can fit and the GPR takes too long. Neural Networks provide a good medium.

I would imagine the random forest would be an okay method too if there is a GPU accelerated version. After checking, there is, but it is single output which is not ideal. But I did figure out that I could pickle models from CuML too. I could probably create a wrapper on top of a single output CuML model to support multioutput (by creating several models) and pickling them to separate model files at the end.

I finished creating the github repository and put some comments and a README in. I think all that’s left are minor edits.

Titan Sims

I haven’t done a whole lot more than what I’ve reported in the previous post. However, I did create a presentation for the APS Eastern Great Lakes Section Conference that I just attended in Marietta, OH. I spent a lot of time researching what TNSA is. In the process, I also started writing up my dissertation by including the Mora model and adding a lot of references to .bib files.