X-ray spectral fitting with Monte Carlo Dropout Neural Networks
We present a novel approach using neural networks to recover X-ray spectral model parameters and quantify uncertainties, balancing accuracy and computational efficiency against traditional frequentist and Bayesian methods. Frequentist techniques often fall into local minima, compromising parameter estimation, while Bayesian methods, though more reliable, suffer from high computational costs. To address these challenges, we apply Monte Carlo Dropout within various neural network architectures trained on simulated spectra generated from a multiparameter emission model convolved with an instrument response. The model parameters are sampled from a predefined prior, and our proof of concept is illustrated using simulated data based on the NICER response matrix for simple emission models with up to five parameters. Our method delivers well-defined posterior distributions comparable to Bayesian inference, achieves accuracy akin to conventional spectral fitting, and is significantly less prone to local minima, thereby reducing the risk of selecting parameter outliers. Moreover, the approach improves computational speed by roughly an order of magnitude compared to traditional Bayesian techniques. This work demonstrates the potential of neural network-based methods as a robust alternative for X-ray spectral analysis, particularly in the context of future astronomical missions expected to generate extensive datasets.