All publications

Robust parameter inference for Taiji via time-frequency contrastive learning and normalizing flows

Tian-Yang Sun, Bo Liang, Ji-Yu Song, Song-Tao Liu, Shang-Jie Jin, He Wang6, Ming-Hui Du, Jing-Fei Zhang, Xin Zhang

  • 6 Corresponding author

PreprintarXiv:2604.138672026

First public:

  • Parameter inference
  • AI for science

Research summary

Transient glitches can distort Bayesian inference for massive black-hole binaries. This preprint combines time-frequency representations, contrastive learning, and conditional normalizing flows to infer source parameters from contaminated Taiji data. A neural glitch generator supports training, while complementary posterior diagnostics assess robustness and distributional agreement across simulated noise conditions.

Extended notes

The framework studies inference under transient contamination and compares alternative encoders and training objectives. Its evaluation supplements coverage checks with a distributional scoring rule.

Sources