A robot and machine learning pipeline helped build a 24.4% efficient perovskite solar cell
The University of Osaka platform filmed perovskite layers forming and found an ideal range for the droplets used to trigger crystals.
In brief
Robots filmed solar film forming, reaching 24.4%
A University of Osaka platform watches perovskite layers form in real time. It found an ideal droplet range. The resulting lead-based cell reached 24.4% efficiency.
An automated platform at the University of Osaka now watches a solar-cell layer form in real time, and it helped produce a lead-based perovskite solar cell with 24.4% power conversion efficiency. That figure is the share of sunlight the cell turns into electricity.
The platform combines robotics, machine learning and rapid optical imaging. The work is described in a paper in ACS Nano titled "Automated Perovskite Film Fabrication with High-Resolution Data Acquisition and Machine Learning Analysis". Reports on it appeared on 2 November 2026.
The problem: a fast, fussy layer
Perovskite is a crystal-like material used as the light-absorbing layer in some solar cells. It starts as a liquid coating and turns into a solid film very quickly. Its quality is highly sensitive to small shifts in temperature, humidity, coating conditions and drying.
Because of that speed, old trial-and-error methods often cannot show which variable really matters.
What the platform does
The system links automated film coating with optical imaging and environmental measurements. It records color changes in the film on a subsecond timescale. It also tracks the motion and impact of droplets within milliseconds.
These droplets are antisolvents, liquids commonly dropped onto the wet film to push it to crystallize. The team focused on them. Machine learning then helped relate many variables to each other.
The result
The analysis identified an optimal range of droplet behavior. Using it to refine the coating process, the team made the 24.4% cell.
Lead author Ryosuke Kado said the aim is to measure what happens while the film forms, instead of relying only on tests after fabrication. Senior author Akinori Saeki said automation allows more data collection than is practical by hand.
Imagine baking a cake whose batter sets in a blink. Filming the oven only after it is done tells you little. Here, the camera runs during the bake.
Why it matters
The team says the approach could help solar research move past empirical tuning toward more reliable, practical cells. The idea is to turn fast events that are hard to see into usable data.
Limits
The sources do not give the cell's size, stability or cost, and do not mention commercial production. Coverage comes from news sites and the ACS Nano listing, not a University of Osaka press release. Reports differ only in emphasis: one stresses machine learning, another the imaging and droplet analysis. All agree on 24.4%.