Emulating the chemistry of the interstellar medium

Non-equilibrium chemistry is essential in astrophysical and cosmological studies, impacting everything from the InterCluster Medium to protoplanetary disks and the InterStellar Medium, key for understanding star formation. These complex systems are analyzed through advanced simulations, now capable of managing trillions of elements due to recent technological advances in computing and algorithms. However, the intricate nature of physical processes in these simulations still presents significant challenges.Describing non-equilibrium chemistry involves solving Ordinary Differential Equations (ODEs) for chemical reactions, heating, cooling, and interactions with gas radiation. The disparity in time scales between chemical and hydrodynamic processes and the exponential increase in reactions with chemical species complicate these equations. This complexity often necessitates the use of expensive implicit solvers and complicates load balancing in large-scale simulations.Recent interest has grown in using scientific machine learning to develop emulators that replace traditional solvers. This thesis explores deep learning approaches to simulate astrophysical chemistry, focusing on Physics Informed Neural Networks (PINNs) and Deep Operator Networks (DeepONet). PINNs, leveraging neural networks’ differentiability and the structural properties of differential equations, offer a new way to approach these problems. DeepONet extends this by not only approximating solutions but also directly mapping initial data to solution spaces.The study demonstrates that PINNs achieve about 10% accuracy compared to traditional solvers and are 100 times faster, reducing computational load significantly. Although they don’t yet match the precision needed to fully replace numerical integrators, their potential is clear. DeepONet, on the other hand, achieves an accuracy of about 1%, making it nearly viable for integration with hydrodynamic models. Despite its high computational efficiency, DeepONet’s reliance on extensive data sets might limit its broader application.In summary, while DeepONet outperforms PINNs in accuracy, the simplicity and cost-effectiveness of PINNs are highly promising for future enhancements and combined applications.