Research map / Time series / Causal and generative time series
TIDMAD: Time Series Dataset for Discovering Dark Matter with AI Denoising
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases TIDMAD, an ultra-long time-series dataset from the ABRACADABRA dark matter detector with hardware-injected signals, a denoising score benchmark, and a dark matter limit benchmark, and benchmarks eight denoising algorithms.
Problem
Dark matter signals (axion-like sinusoids) are buried in detector noise in ultra-long time series at 10 MS/s, and no public ML-ready dataset links denoising to physics limits.
Approach
Pairs noisy SQUID readout with hardware-injected ground-truth sinusoids recorded simultaneously, defines a log-scaled PSD-based denoising score, and automates a community-standard Brazil-band dark matter limit so ML denoising can be translated directly into physics reach.
Key results
- FC Net achieved the best fine denoising score of 6.43 (coarse 6.55), versus 1.00 for raw data and 0.24 for Fourier Averaging.
- WaveNet scored 4.99 fine and PU Net 3.69; all traditional filters (moving average, SG filter) scored below the raw baseline.
- ABRA-TIDMAD denoised dark matter limit improves on the raw limit by 1-2 orders of magnitude and surpasses ABRA-10cm Run 3 at small m_a, using only 1% of Run 3's science data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use deep learning denoisers instead of traditional denoising algorithms | TIDMAD denoising · TIDMAD validation | Fine denoising score: 1.00 -> 6.43 | Helped |
| Use FC Net denoising on the science data instead of leaving the data raw | ABRA-TIDMAD science data · Dark matter limits across different m_a | dark matter limit sensitivity: 1-2 orders of magnitude | Helped |
| Use a smaller Moving Average window size (ablation) | Moving Average · TIDMAD validation | Fine denoising score: 0.86 -> 0.56 | Helped |
| Use smaller window sizes and lower polynomial orders (ablation) | Savitzky-Golay filter · TIDMAD validation | Fine denoising score: 0.95 -> 0.30 | Helped |
| Apply traditional denoising rather than leaving the data raw | Traditional denoising algorithms · TIDMAD validation | Fine denoising score: 1.00 -> 0.24 | Hurt |
| Use coarse rather than fine denoising-score evaluation | Denoising score benchmark · TIDMAD validation | denoising score: 1.00 -> 1.10; 6.43 -> 6.55 | Mixed |
Surprising findings
- Traditional filters (moving average, Savitzky-Golay) with the best tuned parameters never exceed the raw-data baseline score of 1.0 in 60 hyperparameter trials.
- The denoised ABRA-TIDMAD limit surpasses ABRA-10cm Run 3 at small m_a despite the dataset being only 1% of Run 3's size.
Limitations
- Denoised ABRA-TIDMAD limit does not outperform ABRA-10cm Run 3 due to reduced geometric coupling and only 24 hours of data.
- Deep learning models were trained on short segments (4 ms or 2 ms) and frequency-split, limiting generality; the method assumes a null result and does not provide discovery analysis.
- Gaussian noise used to validate the denoising score does not replicate the non-Gaussian real detector noise.
Experimental setup
Datasets: TIDMAD training set TIDMAD validation set TIDMAD science set
Benchmarks: TIDMAD denoising score TIDMAD dark matter limit (Brazil band)
Metrics: denoising score (log of PSD-based SNR ratio) 95% one-sided upper limit on g_aγγ
Baselines: Moving Average Savitzky-Golay filter Fourier Averaging FC Net PU Net Transformer WaveNet RNN Seq2Seq
Builds on: WaveNet U-Net Transformer Focal Loss
Closest papers by meaning
- CosmoBench: A Multiscale, Multiview, Multitask Cosmology Benchmark for Geometric Deep Learning (NeurIPS 2025 · similarity 0.81)
- The Multimodal Universe: Enabling Large-Scale Machine Learning with 100 TB of Astronomical Scientific Data (NeurIPS 2024 · similarity 0.80)
- AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing (NeurIPS 2025 · similarity 0.80)
- Listening to the noise: Blind Denoising with Gibbs Diffusion (ICML 2024 · similarity 0.80)
- Revealing the 3D Cosmic Web through Gravitationally Constrained Neural Fields (ICLR 2025 · similarity 0.80)
- BERTground: A Transformer-Based Model of Background Spectra on the ISS-Based NICER Space Telescope (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.