TIDMAD: A Sandbox for Physics-Informed LLM Agents in Dark Matter Searches

The recently released TIDMAD dataset from the ABRACADABRA experiment provides the first open, ultra-long time-series benchmark for wave-like dark-matter searches. ABRACADABRA is a broadband, magnet-based experiment that uses a sensitive SQUID magnetometer to search for axion and axion-like particles through their coupling to electromagnetism. TIDMAD captures this detector response by providing continuous waveform data, hardware-injected calibration signals, realistic noise conditions, and a fully documented analysis chain leading to likelihood-based exclusion limits. These features create a realistic environment for evaluating analysis methods on detector time-series data.

The presentation will review the design and scientific motivation of TIDMAD and highlight how its modular structure supports physically motivated analysis pipelines for wave-like dark-matter searches. As the dataset provides a fully traceable and physics-grounded workflow, it also serves as a natural sandbox for exploring more ambitious end-to-end AI systems for particle-physics searches. Building on this foundation, the talk will outline initial steps toward physics-informed LLM agents aimed at exploring denoising strategies and characterizing the structure and origin of detector noise.

Controlled perturbations and synthetic noise scenarios within TIDMAD provide a natural setting to probe whether such agents can begin to identify analysis-relevant and physically meaningful behavior, offering an initial view of how physics-informed LLM agents may start to interact with real detector systems.