Google ERA: AI system for scientific research reaches CDC top for hospitalization forecasting, solves an open cosmological problem, and tracks CO2 every 10 minutes

On April 29, 2026, Google Research introduced ERA — Empirical Research Assistance, an internal AI system for scientists that combines LLMs (including Gemini Deep Think) with computational tools to accelerate research. Rather than a hypothetical description, the post offers four concrete examples of results from different domains.

Public health forecasting

ERA generates predictions for hospitalizations from COVID-19, influenza, and RSV. Google’s submissions “rank at or near the top of both leaderboards” for flu and COVID forecasting competitions run by academic institutions, matching or exceeding the performance of CDC tools.

Implication: public health agencies could use similar systems for real-time resource triage.

Cosmology

The team paired ERA with Gemini Deep Think to tackle an unsolved problem in gravitational physics — the energy radiation of cosmic strings. The result: “successfully derived six general solutions and a concise formula for the asymptotic limit”, extending previous partial results limited to specific cases.

This is not a numerical fit — it is a closed-form formula that did not previously exist. It demonstrates that AI assistance can contribute to theoretical physics, not just data analysis.

Climate monitoring

ERA developed a physics-guided neural network that extracts a CO2 signal from GOES-East weather satellites. The model enables unprecedented spatial and temporal resolution: tracking atmospheric CO2 every 10 minutes across entire hemispheres, compared to existing satellite coverage of once every 16 days.

That is ~2,300× more frequent measurement. Implication for climate policy: faster detection of large emission events (wildfires, industry, urban peaks).

Zebrafish neuroscience

Using data on zebrafish neural circuits, ERA discovered “interpretable, mechanistically accurate solutions” linking environmental stimuli to neural responses. This shift — from predictive black-box models toward uncovering actual circuit mechanisms — is epistemologically significant: AI is no longer just a better predictor but a scientific researcher generating mechanistic understanding.

Why does this matter?

ERA demonstrates that AI is moving from NLP/computer vision into the natural sciences as a primary application domain. All four examples are results that could not have been achieved by an LLM alone, requiring combination with domain-specific tools and data. Google Research thereby signals a new competitive arena with DeepMind (AlphaFold, AlphaGeometry), where Google offers a cross-domain AI research assistant rather than specialized per-domain models.