SeisMambaKAN

SeisMambaKAN represents a paradigm shift in seismic phase picking by introducing a novel hybrid architecture that fuses State Space Models (Mamba) with Kolmogorov-Arnold Networks (KAN). The primary challenge in earthquake early warning systems is processing long sequence seismic data with high accuracy and low latency. Traditional Transformers are computationally heavy, while standard CNNs struggle with long range temporal dependencies. To solve this, I designed a lightweight U-Net style architecture (~225k parameters) that utilizes Mamba blocks in the encoder to capture global temporal contexts with linear complexity, and KAN blocks in the decoder for precise, non-linear feature reconstruction. The model is trained on the STEAD dataset using a custom built pipeline featuring WebDataset for efficient streaming and aggressive physics based augmentation (time shifts, noise injection, secondary event blending). The system employs a specialized Multi-Head Loss function combining Focal Loss for detection and Masked Peak MSE for precise P/S phase arrival timing. Currently in the research phase, the model aims for <0.01s phase picking error and >99% detection rates, targeting future deployment as a real time computing solution for Earthquake Early Warning systems.