AI screened 90,000 solar-cell interfaces; the selected material kept over 95% efficiency for 2,000 hours
A peer-reviewed Chinese study combined retrosynthesis prediction, physics calculations and expert rules to select a perovskite-cell interface. The laboratory result is promising, but it is not a field lifetime or mass-manufacturing test.
The short version
- What happened
- Researchers generated more than 90,000 molecular structures from 308 fragments, narrowed them to 300 and ranked them with eight descriptors covering synthesis, electronic behaviour and structure.
- Who it affects
- A more stable, efficient interface could eventually support cheaper or longer-lived perovskite panels. This experiment does not yet show that consumers, installers or grids will receive those benefits.
- What remains unproven
- Laboratory cells rather than commercial-size modules
Keep in mind
Laboratory cells rather than commercial-size modules
2,000-hour accelerated test rather than multi-year outdoor exposure
No factory yield or manufacturing-cost evidence
No lifecycle or supply-chain assessment
Single lead molecule requires independent replication
Source analysis · 1 linked source. Primary analysed source date: 28 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The workflow linked a large digital search to a made-and-tested molecule
The team combined a graph-neural-network retrosynthesis model with density-functional-theory screening and expert chemical rules. More than 90,000 structures assembled from 308 fragments were reduced to 300 candidates, then ranked on eight descriptors—including predicted synthesizability, hole-reorganisation energy and structural features. The highest-ranked molecule, AI-1, was then synthesized rather than remaining a computer proposal.
Perovskite solar cells using AI-1 reached a reported 27.2% power-conversion efficiency. Under maximum-power-point tracking at 65°C, they retained more than 95% of initial efficiency after 2,000 hours. Mechanistic tests linked the result to energetic alignment and sulphur–lead bonding at the buried interface. That connects the search score to measured device behaviour, not just a virtual benchmark.[1]
A durable laboratory cell is not yet a durable solar product
The comparator is the team's other candidate interfaces and device controls, not a commercial module operating for years outdoors. Two thousand hours at elevated temperature is useful accelerated evidence, but it cannot reproduce every combination of moisture, ultraviolet light, thermal cycling, mechanical stress and manufacturing variation. The paper also does not establish the cost or environmental burden of producing AI-1 at scale.
For energy researchers, the practical advance is a traceable route from search space to synthesis and device testing. For households and grid operators, there is no immediate change: the study does not show a saleable panel or lower electricity price. Independent replication, larger-area modules, standardized weathering and manufacturing-yield data would change that assessment.[1]
The distinction that matters
What the AI-guided materials search did
The study moved through prediction, physical screening and laboratory validation.
Generate
Combine 308 molecular fragments
More than 90,000 structures
Rank
Use retrosynthesis, physics and expert descriptors
300 candidates and one lead
Test
Make AI-1 and build cells
27.2% efficiency; >95% retained for 2,000 hours