AI that earns its place in science: UC researchers push accuracy, safety, and real-world impact
Worried about artificial intelligence? You're in good company. Many people see both promise and risk - especially in fields where mistakes carry real costs. Across the University of California, researchers are putting AI to work where it can add value, and pressure-testing it where it can't yet be trusted.
Their stance is practical: use AI to detect patterns humans can't see, and demand proof before handing it the wheel. Here's how that looks in weather forecasting, agriculture, cancer screening, and clean energy.
Why extreme weather is tricky for AI to predict
Modern forecasting ingests billions of daily observations and solves physics at planet scale - a massive computational lift. Ashesh Chattopadhyay at UC Santa Cruz and collaborators built FourCastNet, an AI model trained on decades of reanalysis data to predict near-future weather in seconds instead of hours, using far less compute. It performs on par with traditional methods for many day-to-day conditions.
Major centers have taken note; the European Centre for Medium-Range Weather Forecasts has begun operational use of AI models from several groups, including work in this class of approaches. See ECMWF's announcement.
But extremes expose a weakness. Chattopadhyay's team trained a version of FourCastNet on data that excluded Category 3-5 hurricanes, then tested it on conditions that produce a Category 5. The model detected the storm but capped intensity around Category 2. The physics-based systems did better because the governing equations still encode the possibility space, even without past examples.
Takeaway: AI can be fast and accurate for routine forecasts, yet still miss tail events - the ones decision-makers care about most. The team is now blending climate-informed signals with short-term forecasting pipelines to improve skill on unprecedented storms.
Can AI learn what farmers know by instinct?
Experienced growers can read a leaf like a lab report. For everyone else, UC Davis engineer Alireza Pourreza's group built Leaf Monitor: a handheld system that measures reflectance beyond the visible spectrum and compares it against known signatures for key nutrients like nitrogen, magnesium, and potassium.
Instead of sending samples to a lab and waiting weeks, farmers get a real-time snapshot of plant health in the field. That supports precise irrigation and fertilization, cutting guesswork and reducing inputs where they're not needed. It's a practical step toward site-specific, lower-impact agriculture.
AI in mammography: promise, proof, and the gap between them
Millions of U.S. screening mammograms now include AI support, and the FDA has cleared multiple tools for clinical use. Yet evidence from large, real-world settings is still catching up to the marketing claims.
UCLA radiologist Hannah Milch studied whether an AI tool (Transpara) could flag breast cancers that human readers missed during routine screening. In nearly 185,000 historical exams, the system identified about 30% of interval cancers - cases that presented between screenings. That suggests AI might pull some detections forward by months, when treatment is easier.
The next step is outcomes. Milch and colleagues are co-leading PRISM, a randomized study assigning over 400,000 mammograms to radiologists with or without AI assistance to see if patient results improve. You can track trial details on ClinicalTrials.gov. The goal is simple: confirm safety and effectiveness before broad deployment.
Turning fracking tech into a tool for clean energy
Enhanced geothermal systems aim to deliver round-the-clock, low-carbon electricity by drilling several kilometers into uniformly hot rock, circulating water, and running turbines. Advances in directional drilling make access feasible, but subsurface operations involve complex interactions among stress, heat, fluids, and chemistry - and the risk of induced seismicity is real. (Oklahoma experienced a 900-fold increase in earthquakes from 2008 to 2019 as fluid injection scaled up.)
UC Irvine's Mohammad Javad Abdolhosseini Qomi leads Geophysicist.ai, integrating large language models, physics-based simulators, and site data from across the Western U.S. The objective: better site selection, improved operational envelopes, and clearer risk forecasts for nearby communities.
People-centered progress
Across these projects, the pattern is consistent. Use AI where it saves time, energy, and cost. Keep physics and domain expertise in the loop. Validate on outcomes, not demos. And build with public benefit - safety, equity, and transparency - as non-negotiables.
What to watch next
- Hybrid forecasting that fuses climate signals with short-term AI to improve performance on rare extremes.
- Field-ready spectral tools that move agronomy decisions from batch testing to real-time control.
- Randomized evidence on AI-assisted screening that measures stage shift, recall rates, and downstream outcomes.
- Subsurface AI copilots that combine generative models and physics to reduce seismic risk in geothermal projects.
If you're building internal AI capability for research teams, see curated options by job role at Complete AI Training.
Your membership also unlocks:
