Skip to content

ILLUSTRATIVE · NASA / SVS · NOT ADITYA-L1 DATA

Model cardslightgbm

Model card · candidate detector

LightGBM

Gradient-boosted decision trees.

What is this?

A model card for LightGBM, one of eight detectors evaluated on the frozen Aditya-L1 flare benchmark.

Why should I trust it?

Every figure below is read at build time from the committed benchmark artifact; none is transcribed. The judgements are traceable to the written evaluation protocol and the dataset limitations record.

Where can I verify it?

Section 1

Purpose and intended use

Why it is in the benchmark. The strongest candidate in the benchmark and the one that makes the negative result worth publishing: it is what a practitioner would actually reach for.

Intended use. Ranking minutes by flare likelihood, and testing whether better ranking translates into better operational detection. On this dataset it does not.

Out of scope. Its ranking advantage must not be reported as a detection advantage. At the base rate of this dataset, ranking and alarm burden come apart, and the operational question is settled by the second, not the first.


Section 2

Inputs, outputs, training

Inputs. Fourteen features per minute, all derived from the SoLEXS 1-minute count rate in T1: the log rate, rolling means over 5/15/30/60 minutes, a 15-minute rolling maximum, rolling standard deviations, a background-excess term, two rise rates, and three data-quality terms (GTI fraction, seconds present, partial-minute flag).

Outputs. A calibrated-scale probability per minute; a reliability diagram is published.

Training protocol. Chronological split with a held-out test period beginning at the frozen test start date; no shuffling, so no future minute can inform a past one. The protocol — split, seed, metric set and bootstrap scheme — was fixed in EVALUATION_PROTOCOL.md before any model was fitted.

Operating point. Decision threshold 0.9206, selected on the training period and then frozen.


Section 3

Evaluation

Both tasks, both reported.

M/X NOWCAST · 192,541 held-out minutes · base rate 1.24%

MetricValueWhat it measures
ROC-AUC0.961Ranking quality across all thresholds
PR-AUC0.707The informative curve at a low base rate
Precision0.745Of the minutes flagged, how many were flares
Recall0.628Of the flare minutes, how many were flagged
F10.681Harmonic mean of the two above
MCC0.680Correlation, robust to imbalance
Brier0.016Probability accuracy; undefined for hard classifiers
Balanced accuracy0.813Mean of the two class recalls
Events detected
81 / 82
Event recall 0.988 · 95% CI 0.960–1.000
False alarm runs
79
Of 171 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
ROC-AUC interval
0.949–0.972
95%, day-block bootstrap.

M/X 30-MIN PREDICTION · 192,541 held-out minutes · base rate 1.29%

MetricValueWhat it measures
ROC-AUC0.768Ranking quality across all thresholds
PR-AUC0.103The informative curve at a low base rate
Precision0.179Of the minutes flagged, how many were flares
Recall0.235Of the flare minutes, how many were flagged
F10.203Harmonic mean of the two above
MCC0.193Correlation, robust to imbalance
Brier0.037Probability accuracy; undefined for hard classifiers
Balanced accuracy0.611Mean of the two class recalls
Events detected
36 / 82
Event recall 0.439 · 95% CI 0.256–0.565
False alarm runs
473
Of 540 alarm runs raised. A run, not a minute — the unit an operator actually responds to.
ROC-AUC interval
0.680–0.831
95%, day-block bootstrap.

Confusion matrix · M/X NOWCAST

True negative

189,633

False positive

515

False negative

890

True positive

1,503


Section 4

Feature attribution

Which of the fourteen features the fit relied on.

  • roll_std_600.1406
  • roll_mean_600.1270
  • bg_excess0.1081
  • roll_max_150.0858
  • rise_150.0858
  • roll_std_150.0844
  • log_rate0.0811
  • roll_mean_300.0800
  • rise_50.0788
  • roll_mean_150.0626
  • roll_mean_50.0606
  • gti_fraction0.0053
  • n_seconds_present0.0000
  • q_partial0.0000
Split-based importance, normalised to sum to one. It records where the trees chose to split, not how much each feature changed a prediction.

Attribution is not explanation. These features are strongly collinear — four of them are rolling means of the same series — so credit assigned to one is credit withheld from another, and none of it constitutes a physical account of flare onset.


Section 5

Strengths and failure modes

Strengths

  • The best minute-level ranking of any model evaluated.
  • Emits probabilities that can be checked against observed frequency.

Failure modes

  • Raises substantially more false alarm runs than the threshold at comparable recall — the finding that decides the benchmark.
  • Effective capacity outruns the event count, so apparent gains may not transfer to a different solar-cycle phase.

Section 6

Limitations and ethical considerations

Ethical considerations. This is the model a positive-result incentive would push to the front page. It is published with the result that does not flatter it, which is the point.

Governing limitation clauses

Clause text is not restated here. It is published in full on the dataset card, from the frozen DATASET_LIMITATIONS_FOR_ML.md report, so there is exactly one copy of every caveat.