DELOS: Detecting Shallow Transits in Kepler Photometry Using a Contrastive-Learning Framework

Researchers introduce DELOS, a contrastive-learning framework that detects shallow exoplanet transits in Kepler photometry data with 99.3% validation accuracy. The system outperforms existing detection methods (BLS and TLS) by 15.5% and 11.25% respectively in low signal-to-noise conditions while running 3-80x faster, enabling more efficient searches for terrestrial planets in long-period orbits.

DELOS represents a meaningful advancement in exoplanet detection methodology by applying modern machine learning to a persistent challenge in astronomical data analysis. The framework addresses a genuine limitation in existing transit-detection algorithms: their struggle with shallow signals from distant, long-period planets that produce minimal brightness variations in stellar light curves. By training on 20 million synthetic light curves with realistic noise properties, the researchers created a system that learns to recognize subtle transit signatures without requiring pre-established detection thresholds.

The technical approach combines GPU-accelerated computational efficiency with a one-dimensional convolutional encoder, making it practical for processing massive astronomical datasets. Traditional methods like Box-fitting Least Squares and Transit Least Squares rely on statistical models that become less reliable when signals approach noise floors. DELOS's contrastive learning strategy instead learns what transit-like patterns look like in phase-folded light curves, enabling superior performance exactly where it matters most—detecting Earth-sized planets around Sun-like stars.

The significance extends beyond academic methodology. Current exoplanet surveys like Kepler, K2, and TESS generate petabyte-scale datasets. Accelerating transit detection by 74-80x relative to existing tools fundamentally changes what becomes computationally feasible. The framework's applicability to future missions including PLATO and Earth 2.0 suggests this approach will shape how humanity searches for potentially habitable worlds.

The authors appropriately frame this as methodological validation rather than claiming new discoveries, deferring astrophysical confirmation of candidates. This disciplined approach strengthens credibility. Future work will determine whether DELOS identifies genuinely novel exoplanet candidates in archival data, but the technical validation already demonstrates substantial progress in instrumental sensitivity.

• →DELOS achieves 99.3% validation accuracy on synthetic data while running 3-80x faster than existing exoplanet detection algorithms.
• →The contrastive-learning framework improves precision-recall performance by 15.5% over BLS in low-SNR regimes where shallow transit signals dominate.
• →The system successfully recovered all known shallow intermediate-to-long-period transits in Kepler validation samples without pre-detected threshold events.
• →GPU-accelerated processing and optimized phase-binning make the method practical for petabyte-scale astronomical datasets from current and future surveys.
• →Authors defer astrophysical validation of newly identified candidates to future work, maintaining scientific rigor despite technical breakthroughs.