Energy-efficient GW detection with adaptive expert layers an adaptive expert-driven AI for GW detection

In this paper we propose an Adaptive Hierarchical Mixture of Experts (HMoE) model for energy-efficient, real-time gravitational-wave (GW) detection. The model organises specialized expert subnetworks that focus on different time–frequency characteristics of the GW strain data, and a lightweight gating network that dynamically selects only the most relevant experts for each input. This dynamic expert selection reduces redundant computation and allows the architecture to balance detection accuracy, latency, and power consumption according to signal complexity. To further improve efficiency, the HMoE framework is combined with model-compression techniques, including pruning, quantization, and knowledge distillation, so that most of the inference can be executed with low-precision, low-cost operations while preserving detection performance. The model is trained and evaluated on real LIGO and Virgo datasets from the Gravitational Wave Open Science Center, containing a variety of compact binary coalescence events and realistic non-stationary noise. Compared with convolutional and recurrent neural network baselines, the proposed HMoE achieves higher detection accuracy (98% vs. 85–87%), while reducing average inference latency to 8 ms per sample and average power consumption to about 80 W (versus 130–150 W). These results demonstrate that adaptive expert-driven architectures provide a practical route to sustainable, high-throughput GW detection in power-constrained settings such as remote ground-based observatories, edge devices, and future space-based detectors.