Speaker
Description
We are beginning to encounter computational and inference problems where complex models of atomic physics such as xillver must account for more and more parameters to properly model the absorption and emission lines present within our data. Coupled with the advent of XRISM, spectra tables must become unwieldingly large to compensate for the increase in energy resolution which makes it difficult to load into computers and interpolate between expected spectra. The launch of newAthena in 2037 will make these problems worse as new datasets will require appropriately sophisticated models, which are also becoming more complex as well.
This, along with the fact that our underlying models of this types of data are non-linear, introduce bias in inference of key physical parameters even with lower resolution data and causes parameter spaces to be overly complex and hard to map compared to if you were to use the true analytical model. In this talk, I discuss how surrogate models can serve as alternatives to table models and can help to reduce these problems by interpolating in high dimensional space in a non-linear fashion. I will also discuss the practicalities of actually implementing these techniques and what approaches people should take when first exploring these methods.
