Increasing the Precision of Surrogate Models for Weak Lensing Mass Maps with Flow Matching

@misc{li2026increasing,
  title = {{Increasing the Precision of Surrogate Models for Weak Lensing Mass Maps with Flow Matching}},
  author = {Guangjian Li and Tomasz Kacprzak},
  howpublished = {arXiv},
  year = {2026},
  month = oct,
  doi = {10.48550/arxiv.2605.23114},
  eprint = {2605.23114},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.23114},
}

TY  - UNPB
TI  - Increasing the Precision of Surrogate Models for Weak Lensing Mass Maps with Flow Matching
AU  - Guangjian Li
AU  - Tomasz Kacprzak
PY  - 2026
DA  - 2026/10/27
PB  - arXiv
DO  - 10.48550/arxiv.2605.23114
UR  - https://arxiv.org/abs/2605.23114
AB  - Weak gravitational lensing maps compactly encode the evolution of cosmic large-scale structure and are a key tool for cosmological analyses. Performing inference directly at the map level allows flexible choices of statistics and can increase constraining power. Conventional methods rely solely on N-body simulations and are computationally expensive. Generative machine-learning emulators can accelerate map-level theory prediction. However, existing GAN-based map-level surrogates still have limited statistical fidelity. They can produce over-smoothed maps, may fail to capture the full distribution of generated map sets and can be difficult to train. Continuous normalizing flows trained with flow matching have recently emerged as a powerful class of generative models. We present a residual label-conditional flow matching generative network that conditions explicitly on the matter density Ω_m and clustering amplitude σ_8 for a fixed source redshift distribution n(z). The model learns a continuous probability flow in a residual space from label-specific noise distributions to convergence maps. We evaluate it using pixel and peak statistics, the power spectrum, bispectrum, power-spectrum correlation matrices, and other validation metrics. Compared with the previous GAN benchmark, the proposed method improves the typical fidelity of generated maps from below 10% and below 20% to below 1% and below 5% for basic and higher-order statistics, respectively. The agreement at the level of map distributions is also very good: maps generated from random noise match well the distribution of maps generated with N-body simulations from random initial conditions. This work brings us closer to a practical mass-map emulator that captures the cosmological signal while supporting multiple forms of data analysis.
ER  -