Google Research has described four early real-world applications of Empirical Research Assistance (ERA), an AI system introduced in late 2025 to help scientists generate expert-level empirical software. Rather than acting as a single model, ERA pairs large language models with automated experimentation so researchers can build, test and refine computational models for specific scientific problems. The examples span epidemiology, cosmology, atmospheric monitoring and neuroscience, and together they illustrate how AI is beginning to change not only what science can answer but how the work is done.
Public health: forecasting hospitalisations for flu, COVID-19 and RSV
In an initial preprint, researchers used ERA to predict U.S. hospitalisations for COVID-19 and showed that, applied retrospectively, it could match or outperform established tools from the Centers for Disease Control and Prevention (CDC) and leading research institutions. The team then extended the approach to influenza and respiratory syncytial virus (RSV) and began submitting prospective forecasts in real time each week. When the CDC’s flu forecasting challenge opened for the 2025-26 season, Google began contributing weekly forecasts for every U.S. state across time horizons up to four weeks ahead, and later joined the CDC’s year-round COVID-19 and RSV forecasting hubs. According to public leaderboards maintained by biostatistics professor Nicholas Reich at the University of Massachusetts Amherst, Google’s submissions performed at or near the top during the period covered. The wider significance is that an AI tool able to rival specialist public-health models could broaden access to computational forecasting for more conditions and locations.
Cosmology: gravitational radiation from cosmic strings
Cosmic strings are hypothesised defects in spacetime, thought to have formed in the early universe and predicted to emit gravitational radiation. Calculating the spectrum of that emitted energy has been an open problem, largely because the governing equations contain singularities — points where values approach infinity and conventional methods break down. An earlier effort using a frontier language model had produced only a partial solution, limited to the simplest case of a square loop. Google researchers combined ERA with an advanced reasoning model to systematically explore techniques for handling the singularities, and reported deriving six general solutions along with a concise formula for the asymptotic limit, published in a March preprint. The result is presented as an example of pairing automated empirical software with capable reasoning models to reach precise, novel solutions at the frontier of theoretical physics.
Climate and sustainability: monitoring CO2 with weather satellites
Continuous observation of atmospheric carbon dioxide began at Hawaii’s Mauna Loa Observatory in the late 1950s, producing the well-known Keeling Curve that documents rising global CO2 concentrations. Mapping emissions and the way plants, soils and oceans absorb them requires tracking how CO2 varies across regions and over time. Dedicated space-based sensors such as NASA’s Orbiting Carbon Observatory-2 (OCO-2) make high-precision measurements but cover only a small fraction of the Earth’s surface and revisit each location roughly once every 16 days. Geostationary weather satellites such as GOES East, by contrast, can scan an entire hemisphere every 10 minutes — yet none were designed to measure CO2.
Google researchers used ERA to build a single-pixel, physics-guided neural network that distils a column-averaged CO2 signal from existing GOES East data, combining 16 wavelength bands with lower-troposphere meteorology, solar angle and time of year. After training on the sparse observations from OCO-2 and OCO-3, the model could estimate column-averaged CO2 everywhere, every 10 minutes. Research presented at the International Workshop on Greenhouse Gas Measurements from Space indicated that the model tracks CO2 at high spatial and temporal resolution, with comparisons against further OCO-2 data and the ground-based Total Carbon Column Observing Network used to check that it captures real variability. The work shows how AI can extract additional value from instruments that were never designed for the task.
Neuroscience: uncovering the mechanisms of neural circuits
Although tens of thousands of neurons can now be mapped in living brains, working out the functional circuits that connect them remains a major challenge. Google researchers applied ERA to this problem in real and simulated zebrafish, a common model organism. Using the wiring diagram of a simplified zebrafish brain-and-body simulator — which reveals the connections between cells but not the rules that govern them — ERA proposed circuits linking a visual stimulus to neural activity and motor response. When tested against new stimuli, the AI-hypothesised circuits generalised rather than acting as mere statistical shortcuts, suggesting they captured genuine mechanisms. This builds on earlier work in which AI-developed models outperformed baseline methods at predicting the activity of more than 70,000 neurons in a published zebrafish benchmark.
Limitations and what to watch
These are early, largely self-reported results from Google Research and its academic collaborators, several of them shared as preprints that have not yet completed peer review. Strong performance on retrospective tests or specific benchmarks does not guarantee comparable results in other settings, and forecasting leaderboard positions naturally shift over time. The cosmology and neuroscience findings, while striking, address particular problems and will need independent verification before being treated as settled. As with any AI-assisted science, the value depends on careful validation against real-world data, and outputs are best viewed as accelerants for expert researchers rather than replacements for them. Full details are available from the Google Research blog and the accompanying ERA paper. Readers interested in the techniques behind such systems may also find the guide to systematic prompting a useful companion.
