Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Dynamic Diffusion Schrödinger Bridge in Astrophysical Observational Inversions
Authors: Ye Zhu, Duo Xu, Zhiwei Deng, Jonathan Tan, Olga Russakovsky
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | By investigating its learning process and prediction performance in both physically simulated data and in real observations (the Taurus B213 data), we present two main takeaways. Our comprehensive experiments on both molecular density and magnetic field tasks validate the effectiveness of our proposed method. |
| Researcher Affiliation | Collaboration | Ye Zhu 1,2, Duo Xu3, Zhiwei Deng4, Jonathan C. Tan5,6, Olga Russakovsky1 1Department of Computer Science, Princeton Univeristy, Princeton NJ, USA 2LIX, École Polytechnique, IP Paris, Palaiseau, France 3 Canadian Institute for Theoretical Astrophysics (CITA), University of Toronto, Toronto, Canada 4 Google Deep Mind, Mountain View, CA, USA 5 Department of Astronomy, University of Virginia, Charlottesville VA, USA 6 Department of Space, Earth & Environment, Chalmers University of Technology, Gothenburg, Sweden EMAIL,EMAIL, EMAIL,EMAIL,EMAIL |
| Pseudocode | No | The paper refers to "detailed algorithms in the appendices" (Fig. 2), but upon reviewing the appendices (Sections A, B, C), no explicit pseudocode or algorithm block with structured steps was found. The descriptions are textual or diagrammatic. |
| Open Source Code | Yes | 2Code available at https://github.com/L-Ye Zhu/Astro DSB |
| Open Datasets | Yes | Similar to previous work (Xu et al., 2023b, 2025), our training data are drawn from synthetic magnetohydrodynamics (MHD) simulations. Specifically, we follow the simulation setup from Wu et al. (2020) and Hsu et al. (2023), which are conducted using the MUSCL-Dedner method and HLLD Riemann solver in the adaptive mesh refinement (AMR) code Enzo (Dedner et al., 2002; Wang and Abel, 2009; Bryan et al., 2014). [...] In addition to these synthetic observables, we also evaluate the proposed Astro DSB on real observational data from the Taurus B213 region, using column density maps provided by Palmeirim et al. (2013). |
| Dataset Splits | Yes | For molecular density, we collected data from 7179 simulation trials, split into 5707 training and 1472 testing cases following an 8:2 ratio. For magnetic field strength, we obtained 19100 training and 6370 testing cases. All data are pre-processed by a logarithmic transformation followed by normalization to [0,1], and resized to 128 128 patches. Additionally, we generate 1680 independent tests with different dominate physical process assumptions as the OOD testing cases to evaluate generalization. |
| Hardware Specification | Yes | All models are trained with a batch size of 16 using the Adam W optimizer with a learning rate of 5e-5, running on an AWS cluster with 4 NVIDIA T4 GPUs. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer but does not provide specific version numbers for this optimizer or any other software libraries/dependencies (e.g., Python, PyTorch, TensorFlow) within the text. |
| Experiment Setup | Yes | For the U-Net baseline, we train the model for 20 epochs using Mean Square Error (MSE) loss. For conditional DDPMs, we follow the setup from prior works (Xu et al., 2023a, 2025) and adopt a 1000-step schedule with the standard variational lower bound loss (Ho et al., 2020). Our proposed Astro-DSB is trained with the same time discretization (1000 steps) using the objective defined in Eq. 5. All models are trained with a batch size of 16 using the Adam W optimizer with a learning rate of 5e-5 |