Speaker
Description
Axions are leading dark-matter candidates that may appear in resonant detectors as faint sinusoidal signals buried in detector noise. The TIDMAD benchmark, derived from a real axion dark-matter experiment, provides time-series data sampled at approximately 10 MHz, motivating real-time denoising directly on FPGA-based detector electronics. We present a porting and quantization pipeline for deploying a WaveNet on the TIDMAD dataset. The model is converted to a regression-based architecture, validated across PyTorch and Keras/QKeras implementations, and evaluated using both post-training quantization and quantization-aware training. We find that 16-bit fixed-point quantization preserves the floating-point model’s denoising behavior with negligible numerical degradation. The resulting quantized model is designed for downstream high-level synthesis with frameworks such as hls4ml.