Abstract
Extreme Mass Ratio Inspirals (EMRIs) are key targets for the Laser Interferometer Space Antenna (LISA), yet their analysis is computationally intensive due to complex, long-duration signals and multimodal likelihood surfaces. This paper describes a machine learning framework using Flow Matching Posterior Estimation (FMPE) and Continuous Normalizing Flows (CNFs) to perform parameter estimation in seconds. By leveraging deep learning to navigate the "bumpy" likelihood landscape, the framework offers a scalable path for real-time gravitational wave astronomy, achieving results orders of magnitude faster than traditional Markov Chain Monte Carlo (MCMC) methods.
Introduction
Extreme Mass Ratio Inspirals (EMRIs)—events where a stellar-mass compact object spirals into a supermassive black hole—are essential for mapping spacetime geometry. However, these signals are notoriously difficult to analyze. They persist for months or years, have low signal-to-noise ratios, and exist in a high-dimensional parameter space. Traditional Bayesian methods like MCMC often fail because the likelihood surface is riddled with local maxima, causing samplers to get "stuck" and requiring millions of expensive waveform templates.
Methodology
The authors introduce a deep learning approach based on Flow Matching Posterior Estimation (FMPE), trained on a dataset of 500,000 simulated EMRI samples.
Data Generation and Preprocessing
Waveforms were generated using the FastEMRIWaveforms (FEW) package and processed through FastLISAResponse to simulate the LISA detector's response. To handle the massive data size, signals were converted to the frequency domain via FFT. A max-pooling strategy was then employed to compress the data while retaining the critical physical information necessary for training.
Model Architecture
The architecture consists of a Linear Compression Network ($N_x$) that reduces signals into a 1024-dimensional feature space, followed by a Flow Network of 56 residual blocks. This network utilizes Continuous Normalizing Flows (CNFs), which employ an Ordinary Differential Equation (ODE) solver to transform a simple Gaussian base distribution into the complex target posterior distribution.
The model is trained using "Flow Matching," a technique that is more computationally efficient and stable for ODE-based neural networks than traditional normalizing flow objectives.
Key Results
The FMPE model demonstrates a transformative increase in inference speed without sacrificing statistical validity.
Speed and Robustness
Once trained, the model performs parameter estimation in tens of seconds, compared to the two days required by MCMC benchmarks on similar hardware. Furthermore, the ML model is significantly more robust; while MCMC requires starting points very close to the "truth" to avoid local optima, the FMPE model can successfully scan the entire prior range and identify the correct parameters from scratch.
Statistical Accuracy
The model's performance was validated using P-P plots, confirming that the estimated posteriors are statistically unbiased. In head-to-head tests, the FMPE model successfully recovered injected parameters within 1-sigma or 2-sigma confidence intervals, even in scenarios where randomly initialized MCMC failed completely.
Discussion and Future Work
The results suggest a hybrid strategy: using the ML model to rapidly narrow the parameter space, followed by traditional MCMC for high-precision fine-tuning. Future developments will focus on expanding the model to the full 17-parameter space, incorporating more complex relativistic waveforms, and addressing the "confusion noise" from overlapping signals in real LISA data.
Summary updated