Reduced_spectra_V3.0_FLUX_CALIBRATED

MORPOP spectra flux calibrated to NIRCam. V3.0 consists of some important improvements over V2. Additional long wavelength filters have been successfully introduced for many targets for the first time. These include F335M, F360M, and F444W. At these wavelengths the targets show weak contrast with the nebular background. The filters were excluded in previous flux calibration efforts because the PSF of the targets could not be accurately centroided via the Markov Chain Monte Carlo. Bright and structured nebular emission dragged the MCMC best fitting x and y position away from the target, which automatically triggered rejection of the filter. Upon further inspection of these cases, it was clear that the PSF should be detectable, and that using purely an MCMC inference approach was not the best technique. Rather, a convolutional neural network (CNN) more closely mimics the human eye, recognising that a faint PSF is present inside of complex nebulosity, because the distinctive shape of the PSF is not easily mistaken for nebular emission. This is a detail MCMC is blind to. A CNN was trained on artifical saturated PSFs injected into the ONC NIRCam mosaics. The centroiding accuracy improved dramatically after this training, with almost all filters having sub-pixel accuracy. The filters F335M, F360M, and F444W were routinely off by >5 pixels using the pure MCMC approach. The Gaussian Process background subtraction moved from 1D to 2D. Originally we used 1D models fit to each row and each column of the cutout containing the star. For each pixel, the background model consisted of an interpolation of the associated row and column model. This worked reasonably well, but created unphysical 'grid-like' patterns in some cases. We now use a proper 2D Gaussian Process model. This approach had been tested earlier, but was unstable close to the star, where the model would explode in predicted flux, leading to an oversubtraction of the stellar light. This was ultimately resolved by adjusting the length scales to ignore the sharp/high frequency scales of the PSF. It was also computationally expensive to fit the ~1000 cutouts for MORPOP targets, with computation time scaling with O(n^3) where n is the number of pixels in our 200x200 pixel cutouts. The 2D GP has been fully adopted now after running a set of simulations in which model PSFs were injected into the NIRCam mosaic, and the background was modelled and compared to the real background which is known ahead of time. The GP 2D model performs equally well independent of whether the 200x200 cutout was fully sampled, or subsampled down to a factor of 45! (The GP only sees 1 in every 45 pixels ). This dramatically reduced computation time without affecting accuracy. Generally speaking, the 2D GP background model is accurate to ~a few percent in the ONC. Both of these changed were implemented and the MORPOP spectra were recalibrated. Both the PSF/Aperture photometry ratio and PSF/Synthetic photometry ratio distributions tightened from ~20% to <15%. Tthe bias in the PSF/Aperture ratio finally disappeared, shifting from ~0.9 to ~1.0. Both distributions also draw on significantly more data-points. PSF/Aperture doubled in size thanks to these improvements!