Implementation of an Autoencoder Neural Network to Nuclear Reaction Kinetics

Supernovae are the most colossal and awe-inspiring events that mark the end of some stars' lives, playing a crucial role in the evolution of galaxies and the synthesis of elements essential for life as we know it. During the collapse of massive stars, the extreme conditions can induce fusion reactions, creating new atomic nuclei in a process called nucleosynthesis. The creation of heavier elements is thought to be predominantly produced by supernovae. The work here holds significance for studying both core-collapse and thermonuclear supernovae. Astrophysicists seek to unravel the complexities of supernovae nucleosynthesis through simulations that use nuclear reaction kinetics. Simulating NRK for supernovae nucleosynthesis becomes increasingly impractical as the network size grows. The primary challenge lies in accelerating these simulations to handle the expanding network size effectively. One approach to address this is the use of machine learning to reduce the time and computational demand of these simulations for real-world applications. Specifically for this work, we implement an Auto-Encoder (AE) for a reaction network in nuclear reaction kinetics using the code XNet. The AE is first trained to find a lower-dimensional representation of the data, and then it reconstructs the data using this lower-dimensional representation. The data set used to train the AE is comprised of molar abundances from a 14-species nuclear network evolved using XNet. In this work, we explore the application of AEs as a means to achieve a lower-dimensional representation of nuclear composition data within XNet, offering promising advancements in both accuracy and computational efficiency.