Cosmo-SPINN: Fuzzy Dark Matter Simulations with Physics-Informed Generative Networks
Abstract
Generative machine learning models have recently emerged as powerful tools for producing cosmological simulations. However, many existing emulators do not explicitly enforce the underlying physical dynamics governing cosmological evolution, often leading to artifacts and poor adherence to the evolution equations. In this work, we present a physics-informed generative U-Net framework for fuzzy dark matter (FDM) that addresses two complementary tasks: (i) the evolution of cosmological fields from initial conditions to an arbitrary cosmological scale factor and (ii) the super-resolution of FDM simulations at a specified cosmological scale factor. Our model incorporates a physics-informed loss function that explicitly enforces consistency with the underlying Schrödinger-Poisson (SP) dynamics during training. For the evolution task, we find that the inclusion of this physics-based loss significantly improves the quality of the predicted simulations, even when only a small amount of training data is available. Using only 20% of the training data, the model accurately reproduces the target simulations in a 1 h−1 Mpc box. Furthermore, the framework generalizes effectively across previously unseen realizations of the initial conditions. For the super-resolution task, we present, for the first time, a generative super-resolution model trained on FDM simulations obtained by solving the full SP equations, considering both single and multiple realizations of the initial conditions and analyzing the role of the physics-informed loss in each case. Our approach enables modern generative modeling of cosmological simulations while maintaining physical consistency and substantially reducing generative artifacts.
Versions
Related
- Searching for signatures of self-interacting dark matter in halos from full physics simulations: From 3D structure to projected observables
- Reconstructing the generalized Barrow holographic dark energy with physics-informed neural networks
- Self-Interacting Dark Matter: From Particle Physics Theories to Structure Formation Simulations
- Boosted Physics-Informed Neural Networks for Solving Homogeneous and Non-Homogeneous Lane-Emden Type Equations
- C3NN-SBI: Learning hierarchies of N -point statistics from cosmological fields with physics-informed neural networks
BibTeX
@misc{mishra2026cosmospinn,
title = {{Cosmo-SPINN: Fuzzy Dark Matter Simulations with Physics-Informed Generative Networks}},
author = {Ashutosh Kumar Mishra and Emma Tolley and Nicolas Cerardi},
howpublished = {arXiv},
year = {2026},
month = jul,
doi = {10.48550/arxiv.2607.28604},
eprint = {2607.28604},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.28604},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Cosmo-SPINN: Fuzzy Dark Matter Simulations with Physics-Informed Generative Networks
AU - Ashutosh Kumar Mishra
AU - Emma Tolley
AU - Nicolas Cerardi
PY - 2026
DA - 2026/07/30
PB - arXiv
DO - 10.48550/arxiv.2607.28604
UR - https://arxiv.org/abs/2607.28604
AB - Generative machine learning models have recently emerged as powerful tools for producing cosmological simulations. However, many existing emulators do not explicitly enforce the underlying physical dynamics governing cosmological evolution, often leading to artifacts and poor adherence to the evolution equations. In this work, we present a physics-informed generative U-Net framework for fuzzy dark matter (FDM) that addresses two complementary tasks: (i) the evolution of cosmological fields from initial conditions to an arbitrary cosmological scale factor and (ii) the super-resolution of FDM simulations at a specified cosmological scale factor. Our model incorporates a physics-informed loss function that explicitly enforces consistency with the underlying Schrödinger-Poisson (SP) dynamics during training. For the evolution task, we find that the inclusion of this physics-based loss significantly improves the quality of the predicted simulations, even when only a small amount of training data is available. Using only 20% of the training data, the model accurately reproduces the target simulations in a 1 h^−1 Mpc box. Furthermore, the framework generalizes effectively across previously unseen realizations of the initial conditions. For the super-resolution task, we present, for the first time, a generative super-resolution model trained on FDM simulations obtained by solving the full SP equations, considering both single and multiple realizations of the initial conditions and analyzing the role of the physics-informed loss in each case. Our approach enables modern generative modeling of cosmological simulations while maintaining physical consistency and substantially reducing generative artifacts.
ER -