Detection and Prediction of Future Massive Black Hole Mergers with Machine Learning and Truncated Waveforms

Abstract

We present a novel machine learning framework tailored to detect massive black hole binaries observed by spaceborne gravitational wave detectors like the Laser Interferometer Space Antenna (LISA) and predict their future merger times. The detection is performed via convolutional neural networks that analyze time-evolving Time-Delay Interferometry (TDI) spectrograms and utilize variations in signal power to trigger alerts. The prediction of future merger times is accomplished with reinforcement learning. Here, the proposed algorithm dynamically refines time-to-merger predictions by assimilating new data as it becomes available. Deep Q-learning serves as the core technique of the approach, utilizing a neural network to estimate Q-values throughout the observational state space. To enhance robust learning in a noisy environment, we integrate an actor-critic mechanism that segregates action proposals from their evaluation, thus harnessing the advantages of policy-based and value-based learning paradigms. We leverage merger estimation obtained via template matching with truncated waveforms to generate rewards for the reinforcement learning agent. These estimations come with uncertainties, which magnify as the merger event stretches further into the future. The reinforcement learning setup is shown to adapt to these uncertainties by employing a policy fine-tuning approach, ensuring the reliability of predictions despite the varying degrees of template-matching precision. The algorithm denotes a first step toward a low-latency model that provides early warnings for impending transient phenomena. By delivering timely and accurate forecasts of merger events, the framework supports the coordination of gravitational wave observations with accompanied electromagnetic counterparts, thus enhancing the prospects of multi-messenger astronomy. We use the LISA Spritz data challenge for validation.