Google neural network finds an eighth planet, Kepler-90i

AI system

System: Deep convolutional neural network for Kepler transit-signal classification (Shallue and Vanderburg)
Model: Not assessed
Release: Not assessed
Application developer: Google (Christopher Shallue, Google Brain) with Andrew Vanderburg (University of Texas at Austin)
Role: Not assessed
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Large labeled set of vetted transit signals and false positives. Convolutional network trained on folded light curves. Application to systems with known planets to look for weaker signals. Statistical validation of high-confidence candidates.
Human oversight: Not assessed
Relation to AI: The neural network ranked the weak Kepler-90i transit signal above false positives in systems that earlier automated and human searches had already examined. The authors then statistically validated both new planets. The AI did the identification. Validation used conventional methods.
AI contribution: The neural network ranked the weak Kepler-90i transit signal above false positives in systems that earlier automated and human searches had already examined. The authors then statistically validated both new planets. The AI did the identification. Validation used conventional methods.

Outcome identifiers: Not assessed
Benefit: a rocky planet about 30 percent larger than Earth with a 14.4-day orbit
Category: Improved Knowledge Discovery
Status: Not assessed
Beneficiary groups: Organizations. Society.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: a rocky planet about 30 percent larger than Earth with a 14.4-day orbit
Comparison: the automated tests and human vetting previously used to verify Kepler signals
Results: Reference: the weakest signals often are missed using these methods. Result: the neural network correctly identified true planets and false positives 96 percent of the time.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: Google (Christopher Shallue, Google Brain) with Andrew Vanderburg (University of Texas at Austin). Operator: Shallue and Vanderburg, applied to NASA Kepler archival data.
Jurisdictions: Developer: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: The host star's light curve must contain the transit at detectable but sub-threshold signal-to-noise. Validation must rule out astrophysical false positives independently of the classifier.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: 25 March 2019. The same group adapted the network as AstroNet-K2 for the K2 mission and used it to identify and validate two previously unknown exoplanets.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: Organizations. Society.
Depth: Organizations: 2. Society: 2.
Reasons: Organizations: NASA's Paul Hertz said that discoveries were lurking in archived Kepler data awaiting the right tool. The archive yielded two newly validated planets without new telescope observations by the authors. Their validation used Kepler data and follow-up observations that other observers had already uploaded to the Kepler Community Follow-up Program. Society: Astronomy gained a validated eighth planet and the first exoplanetary system known to have eight planets, tying the Solar System. This is a knowledge gain for a specialist community rather than a population-level benefit.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: The makers' claim

Kind: Not assessed
Related entries: Not assessed
Field: Astronomy / exoplanets. Academic research.
Purpose: Built to classify Kepler transit signals as planets or false positives and to recover planets missed by earlier searches. This benefit was that purpose.
Dates: Date basis unassessed: 2017 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 4a57491a95cc…