Abstract
Determining the absolute mass scale of neutrinos is a fundamental challenge in particle physics and cosmology. While massive neutrinos suppress the growth of structure on small scales, modeling this effect traditionally requires computationally intensive N-body simulations. We introduce the Deep Density Displacement Model (D3M), a 3D convolutional neural network designed to map massless neutrino simulations to their massive neutrino counterparts. D3M produces simulations 10,000 times faster than traditional N-body codes while maintaining high accuracy across linear and non-linear scales, offering a transformative tool for analyzing data from next-generation cosmological surveys.
Introduction
Neutrinos are among the most abundant particles in the universe, yet their exact masses remain unknown. In cosmology, massive neutrinos act as "hot dark matter," moving at high velocities that prevent them from clustering on small scales. This results in a characteristic suppression of the matter distribution, which can be used to constrain the sum of neutrino masses ($M_\nu$).
However, extracting these constraints from upcoming surveys like Euclid or the Vera C. Rubin Observatory (LSST) requires a vast library of theoretical templates. Traditional N-body simulations, which solve the gravitational equations for billions of particles, are the gold standard for accuracy but are prohibitively slow, often requiring hundreds of CPU hours per realization. This creates a significant bottleneck for exploring the high-dimensional cosmological parameter space.
Methodology: The Deep Density Displacement Model (D3M)
To overcome this bottleneck, we developed D3M, a specialized 3D convolutional neural network based on the U-Net architecture. Unlike previous approaches that attempt to predict density fields directly, D3M learns to predict the displacement field—the vector representing the movement of particles from their initial grid positions to their final state at redshift $z=0$.
The model utilizes a 3D U-Net with 15 layers, featuring a contracting path to capture large-scale gravitational context and an expansive path to reconstruct high-resolution spatial details. It was trained on the HADES simulation suite, mapping the displacement fields of simulations with massless neutrinos to those with $M_\nu = 0.15$ eV. By focusing on the "residual" displacement caused by neutrinos, the network efficiently learns the subtle physical perturbations they induce.
Results and Analysis
D3M demonstrates remarkable fidelity in capturing the complex, non-linear structures of the cosmic web. Visual inspections of the density fields and residual maps show that D3M significantly reduces the error compared to simply using a massless neutrino simulation as a proxy.
The statistical accuracy of the model is further validated through the matter power spectrum $P(k)$, which measures the clustering strength at different scales. D3M maintains approximately 6% accuracy down to $k = 0.7 \, h/\text{Mpc}$, significantly outperforming linear approximations like the Zel’dovich Approximation (ZA) on non-linear scales.
Beyond the power spectrum, the model accurately reproduces higher-order statistics, including the bispectrum (with a mean relative residual of only 0.4%) and void abundances. This indicates that the CNN has learned the underlying structural physics of neutrino-induced suppression rather than merely performing image-to-image translation.
Conclusion and Future Work
The D3M model represents a significant leap in cosmological simulation efficiency, providing a 10,000-fold speedup over N-body methods. While current performance is excellent on large and intermediate scales, accuracy begins to decline at very small scales ($k > 0.7 \, h/\text{Mpc}$) where highly non-linear processes dominate.
Future developments will focus on training with higher-resolution datasets, such as the QUIJOTE simulations, and extending the model to be "universal"—capable of taking any neutrino mass as an input parameter. This approach paves the way for real-time theoretical predictions in the era of "big data" cosmology.
Summary updated