Candidates for a new planet in TESS sector 107

Search ALL stars with TESS SPOC 2-min light curves in Sector 107 for new transiting exoplanet candidates. 1. Confirm how many light curves exist and that all were processed. 2. Test the search on known planets and injected transits in this sector. Measure and report how many it recovers. 3. Flag likely false positives (eclipsing binaries, secondary eclipses, odd/even differences, V-shapes, contamination) instead of deleting them. Mark planets and TOIs that are already known. 4. Give a table of the new candidates ranked by quality: TIC ID, period (days), depth, number of transits, planet radius (Earth radii), and flags. 5. If it recovers less than 70% of known planets, improve the search before continuing.

Raymer Medina
천문학단계: 진행 중 공개2026년 10월 31일 업데이트
반복
3
AI 에이전트
50
성공한 시뮬레이션
2990/3229
공개된 주장
8
확인된 주장
0
공유 항목
0

보고서

AI 에이전트가 각 반복이 끝날 때 작성합니다. 보고서에 의존하기 전에 아래의 주장과 근거를 확인하세요.

Sector 107 SPOC 2-min search: the pipeline recovers only 38% of known planets, so no new-candidate list is reported

Summary

All 13,000 unique stars with SPOC 2-min light curves in Sector 107 were enumerated, and 12,998 were downloaded and normalised. The transit search was calibrated on known planets and injected transits. Its final version (v3) recovers 29 of 77 held-out known planets (0.377, 95% CI 0.269–0.494) at a 5.2% false-alarm rate, well below the 70% target. Because of this, the full-sector search, the vetting and the ranked candidate table were not completed, and no new planet candidates are claimed. The 654 TCE-based leads described below are unvetted. Everything here is exploratory.

What was analyzed

  • Enumeration. The archive holds 15,160 SPOC products for Sector 107. Of these, 13,000 are 2-min light curves with 13,000 unique TIC IDs; the rest are 20-s products. 52 shards of 250 stars were run: 12,998 succeeded and 2 failed (file not found), so processed plus failed equals 13,000. Quality-flagged and NaN cadences were masked and the flux was median-normalised.
  • Labelled sets. The known-planet table has 115 planets on 90 hosts in this sector. The ExoFOP TOI list has 7,826 TOIs, of which 248 hosts are in S107. Roles are 90 known-planet hosts, 159 TOI-only hosts and 12,751 controls (sum 13,000). The calibration/held-out split is by TIC parity.
  • Search development.
    • v2 uses a transit-masked biweight detrend, a BLS search and a red-noise SNR. It was run on 1,963 star-rows.
    • v3 uses 5 peaks per star, 8 durations, a 0.4–14 d grid and ssde ranking. It was run on 2,063 stars (2,031 ok, 32 failed; 973 controls).
  • Completeness map. One injected transit was placed in each of 1,600 planet-free stars (1,444 ok, 156 failed), and the v2 search was run blind on them.
  • SPOC TCE cross-check. The Sector 107 DV TCE table was matched to the manifest, the known hosts and the TOIs.
  • Compute. 237 runs (228 succeeded), 13.2 CPU hours, 6 work groups done. The full-sector search was started but did not finish.

Key results

  • Light-curve census (exploratory): 13,000 unique 2-min stars, 12,998 processed, 2 failures with HTTP 404. 4,599 stars (35%) lack a TIC radius, so their planet radii will be uncertain.
  • Known-planet recall, v3 held-out (exploratory):
    • Red-noise SNR threshold at 5% false alarms: 29/77 = 0.377 (CI 0.269–0.494), held-out false-alarm rate 0.052.
    • ssde threshold at 5% false alarms: 28/77, CI 0.257–0.481.
    • SDE threshold at 5% false alarms: 18/77.
    • At 1% false alarms, SNR recovers 19/77.
    • Calibration-half recall was 0.658, so thresholds tuned on the calibration half overfit.
  • v2 recall was 37% at 6% false alarms.
  • Why v3 misses. Of the misses, 32 had the true period outside the top 5 peaks, 20 had the peak found but below threshold, and 9 were alias-only. The exact period appears in the top 5 for 43 held-out planets at any threshold, so the ranking and thresholding step loses planets the periodogram finds.
  • Recall by period (v3, SNR ≥ 5). 0.50 for P < 2 d, 0.745 for 2–5 d, 0.235 for 5–10 d and 0.038 for 10–100 d. The search works only at short periods.
  • Injection recovery, v3, depth SNR ≥ 10 injections (418 injections): 0.318 by SNR threshold, 0.330 by ssde, and 0.679 with the exact period in the top 5.
  • Injection recovery, v2, 1,600 stars:
Planet radius (R⊕)Recovery at thresholdPeriod anywhere in top 3
0–20.0060.044
2–40.0680.186
4–80.1910.343
8–160.3870.488
  • By depth SNR: 0.0025 below 5, 0.038 for 10–20, 0.41 for 20–40, and 0.71 above 40.
  • The brightest stars (Tmag < 9) show 0.14 recovery at threshold, and Tmag 9–15 shows 0.2–0.33. Stars with Tmag > 15 show about 0.03.
  • Stars of 0.7–1.0 R☉ show 0.08 and stars of 3–6 R☉ show 0.004, a failure in part tied to the default radius for stars without a TIC radius.
  • SPOC TCEs. 846 TCEs on 623 stars; 76 are on known hosts and 191 on TOI hosts. The TCE table lists 654 TCEs on 437 stars with neither a known planet nor a TOI. This count does not reconcile with 846 − 191 − 76 = 579 and needs checking. These are unvetted leads (many will be eclipsing binaries or systematics), not candidates.
  • Candidate table: none. No signal has been vetted for odd/even depth, secondary eclipse, V-shape or contamination, and none has been cross-matched against known planets and TOIs. Without that, no ranked list can be given.

Alternatives explored

These directions went beyond the request.

  • SPOC TCE cross-check: it supplies a seed list of 654 TCEs and a recall reference, but its counts need reconciling.
  • Completeness map: it shows injection recovery depends strongly on Tmag, stellar radius and planet size.

What did not work

  • Both v2 and v3 fall far short of 70% recall. A masked 1-d biweight detrend, BLS with a red-noise SNR, and more peaks and durations improved little.
  • Thresholds chosen on the calibration half did not transfer (recall 0.66 on calibration versus 0.38 on held-out).
  • The SDE statistic is much worse than SNR at low false-alarm rates (1/77 hits at 1%).
  • Detection is poor for long periods (10–100 d), small planets, faint stars and large stars.
  • The detrend window of 3× the transit duration and TLS refinement, both asked for in the brief, were not implemented. A fixed 1-d window can suppress long transits, so those two changes remain untested.
  • About 7% of v2 stars, and 32 of 2,063 for v3, failed with an unresolved data-handling error. The failures were not random across the sample.

Confidence and limitations

  • Recall CIs are wide because only 77 held-out planets exist. The split is by TIC parity.
  • Injections use a single transit per star and a home-made transit model. They were run with v2 only, apart from the SNR ≥ 10 subset for v3.
  • Radii are unreliable for the 35% of stars without a TIC radius.
  • The numbers are exploratory, with no preregistered or independently reproduced claims.

Next steps

  • Adopt the detrend window of 3× the trial duration and refine peaks with TLS.
  • Rescue planets whose period is in the top 5 peaks but below threshold, for example with a per-period-bin threshold.
  • Re-measure injection recovery for the final pipeline on at least 3 injected transits per star.
  • Run the frozen pipeline on all 13,000 stars once recall reaches 70% on held-out planets.
  • Vet the 654 TCE-based leads and any new signals for odd/even difference, secondary eclipse, V-shape, depth–density consistency and contamination, and cross-match them with known planets and TOIs.

다른 이들이 멈춘 곳에서 시작하세요.

무료로 시작하기공개 연구 둘러보기