AI and GW

Gravitational-wave signals are extremely small and are buried in noise. We also do not know when they will arrive. Only after narrowing down the time and frequency ranges and applying filters can we finally begin to see them, and even then it is not guaranteed.

The most effective way to filter out noise is known as matched filtering, in which the telescope output is multiplied by a waveform that is exactly the same as the expected gravitational-wave signal. For example, gravitational waves produced by a binary black-hole system have a waveform called a chirp signal, in which both the frequency and amplitude increase with time. In this case, the best approach is to change the filter frequency in step with time. However, the waveform (template) depends on factors such as the masses of the black holes, and the arrival time at the detector is unknown, so all possibilities must be searched exhaustively. Even when ignoring details such as spin, this still requires hundreds of thousands of templates.

Recently, artificial intelligence (AI) has begun to play a role in gravitational-wave observations. Unlike humans, AI evaluates data using different criteria and can combine information in nonlinear ways—not only by weighting different parts of the data, but also by multiplying or dividing one part by another. Although its accuracy does not yet surpass matched filtering, AI can find signals more quickly. In practice, matched-filter searches ignore details such as the directions of black-hole spins in order to reduce analysis time. If AI searches include these effects, it may even discover signals in past data that were previously missed.

However, adding more conditions increases the time required to train the AI. While detection is fast after training, an extremely long training time makes practical use difficult. Training also requires gravitational-wave signal data prepared by hand. Because calculating waveforms with numerical-relativity codes takes a long time, collecting many different signal examples is itself challenging. For these reasons, it is not as simple as pressing a button and immediately having AI find gravitational waves.