Google AI team helps NASA identify a new planet in the Kepler-90 star system, found with the aid of a neural network using data from the Kepler Space Telescope
Google has previously discovered lost tribes, missing ships and even an forgotten forest. But now it has also found two entire planets.
Context & Ripple Effects
By late 2017, Google's neural networks had already proven themselves on perception problems closer to home — the PlaNet model that outperformed humans at geolocating images a year earlier showed the same pattern-recognition stack could be pointed at raw sensor data rather than consumer products. The Kepler-90 result extends that playbook off Earth: instead of building a new telescope, Google's AI team and NASA trained a network to re-scan archival Kepler Space Telescope data for transit signals astronomers had missed.
The significance is the division of labor it establishes — NASA supplies the instrument and the labeled examples, Google supplies the model — and the coverage since shows both sides kept investing in it: NASA's AI-driven hunt for fresh Martian craters in orbital imagery applied the same archive-mining method inside our own solar system, while Google's later Gemini for Science tools formalize the researcher-facing side of what began as one-off collaborations.
First-order effects
- NASA gains a validated method for extracting additional discoveries from Kepler's already-collected data, turning a finished mission's archive into an active detection pipeline at near-zero marginal cost.
- Google's AI team converts a research demo into scientific standing — a peer-credible planet discovery that differentiates its AI work from consumer-product rivals.
Second-order effects
- Other observatories and space agencies with large under-explored image archives now face pressure to partner with machine-learning teams or build in-house capability, since the bottleneck shifts from telescope time to analysis capacity.
- Tech companies gain a new currency in government-science relationships — compute and models — positioning them as indispensable infrastructure partners rather than vendors.
Third-order effects
- If the pattern holds, discovery itself migrates toward AI-screened archives: hypotheses generated by models combing existing datasets first, with new instruments commissioned to confirm rather than explore — a structural change in how astronomy budgets are justified.
- The collaboration template points toward institutionalized AI-for-science programs inside agencies like NASA, with corporate labs embedded in the discovery workflow rather than consulted ad hoc.
The trend: Scientific discovery is shifting from instrument-limited to analysis-limited, as tech-company neural networks turn completed missions' data archives into ongoing sources of new findings.