3D-PDR Orion dataset and NeuralPDR: Neural Differential Equations for Photodissociation Regions

TL;DR

The paper tackles the computational bottleneck of simulating photodissociation regions by introducing the 3D-PDR Orion dataset, comprising 8192 1D PDR models inspired by the Orion Bar and spanning key physical parameters. It proposes neural differential equation emulators, specifically Augmented Neural Ordinary Differential Equations (ANODEs) with auxiliary parameters, and compares direct evolve and latent encoder–evolve–decode architectures. The results show that well-regularized evolve ANODEs achieve high-fidelity replication of molecular abundances and temperatures at a fraction of the computational cost, with inference times suitable for coupling into 3D simulations. This work provides a fast, scalable surrogate framework for astrochemistry along lines of sight, enabling higher-resolution PDR studies and future extensions to line traces.

Abstract

We present a novel dataset of simulations of the photodissociation region (PDR) in the Orion Bar and provide benchmarks of emulators for the dataset. Numerical models of PDRs are computationally expensive since the modeling of these changing regions requires resolving the thermal balance and chemical composition along a line-of-sight into an interstellar cloud. This often makes it a bottleneck for 3D simulations of these regions. In this work, we provide a dataset of 8192 models with different initial conditions simulated with 3D-PDR. We then benchmark different architectures, focusing on Augmented Neural Ordinary Differential Equation (ANODE) based models (Code be found at https://github.com/uclchem/neuralpdr). Obtaining fast and robust emulators that can be included as preconditioners of classical codes or full emulators into 3D simulations of PDRs.

Figures (5)

• Figure 1: Comparison between different NeuralPDR configurations, the y-axis is the index of the visual extinction. The loss of the a,b,c and eevolve models largely overlap.
• Figure 2: The result for one sample as inferred by the (e-a) emulator as a function of the log-$A_V$. The MSE in the feature space for this sample is listed in parenthesis behind each feature.
• Figure 3: The distribution of visual extinction index versus the actual log values of the index. The 5 red lines are 5 individual samples, with the 2d histogram the distribution.
• Figure 4: A comparison of the validation for the losses, with a moving average of 10 steps. The final loss is denoted in the legend behind each architecture.
• Figure 5: The MSEs for each individual molecule and all other features, for all different architectures.