Model dispersion with PRISM

CI: – Ellert Van Der Velden

This ADACS project aims to increase the potential of the PRISM pipeline, a Python package that I am developing for my PhD, by utilizing the power of GPUs where appropriate. PRISM is based on the idea of emulation and history matching, allowing for an approximation of a model to be created by using polynomial functions. This involves large numbers of repetitive calculations to sift through combinations. Once the model has been mapped, or emulated, it can be compared with data allowing significantly faster scanning of a model parameter space for ‘regions of interest’ than MCMC methods; this is particularly valuable for complex models that have significant runtimes. Although this idea has been used successfully several times already, the algorithms/frameworks/codes for it have never been publicly released. This makes PRISM the first publicly available emulation framework which will allow the national and, indeed international community , to use it for their own research. PRISM has been created with a modular and entirely generalised design, allowing any model to be connected to it (unlike the previously built frameworks, which were specialized). Therefore any astronomy field, be it observational or theoretical, that use a model codebase (e.g. stellar population synthesis, CLOUDY ionisation tables or semi-analytic models) can be significantly improved by fully exploring their parameter space and highlighting regions of interest, far more efficiently than
the current MCMC standard.