PV for AI, AI for PV – Part 2 of 4
In our first post, we explored how solar can help meet AI’s growing electricity demand. The reverse relationship could be equally significant: AI can accelerate the discovery and development of better solar cells. But faster discovery must translate into cells that last, are affordable and can be manufactured reliably at scale.
That is a demanding research challenge given the sheer number of possibilities: Material compositions, layer thicknesses, interfaces and processing conditions interact. A change that improves efficiency may compromise stability or make fabrication harder to reproduce. The likely next-gen materials or PV cell architecture like Perovskites and tandem cells expand the possibilities… and make navigating these trade-offs even more important.
AI can help researchers decide which experiment to run next.
Instead of testing every combination, models learn from simulations and experimental results to identify promising candidates, flag uncertainty and guide further testing.
The results are already tangible. Researchers at KIT and Helmholtz Institute Erlangen-Nürnberg combined machine learning with automated synthesis to explore a space of approximately one million potential molecules for perovskite solar cells. Around 150 targeted experiments identified materials that helped achieve 26.2% cell efficiency, approximately two percentage points above their reference device.
The opportunity also extends to understanding cell structures. Digital twins combining AI with device physics can connect measurements to otherwise hidden electrical properties and reveal where performance is being lost. Fraunhofer ISE is developing this approach to support solar-cell analysis and targeted optimisation. For researchers, this provides a stronger basis for deciding which design changes deserve laboratory testing.
Connecting these capabilities creates the prospect of an automated laboratory that learns continuously: AI proposes a candidate, robots prepare the sample, instruments measure its behaviour, and the results inform the next experiment. A 2026 Nature study demonstrated such a loop for perovskites, combining machine-learning-guided molecular discovery with automated device fabrication and reporting substantially improved reproducibility compare with manual fabrication.
The next challenge is building a shared data foundation. Reliable AI needs high-quality, representative data linking material composition and processing conditions to cell performance and stability not forgetting all unsuccessful experiments. Research laboratories contribute exploratory breadth; pilot lines and manufacturers add evidence on reproducibility and scale. Connecting these complementary datasets through common standards and agreements protecting industrial know-how could accelerate learning across the sector. Open resources such as the Perovskite Database Project provide a starting point. The ambition should be a shared PV knowledge base, inspired by AlphaFold’s accessibility, where predictions are clearly distinguished from experimentally validated results.
The strategic objective should be more validated knowledge per experiment… and a shorter path to commercially useful PV cells. A simulated optimum still needs experimental proof; a laboratory record still needs long-term validation.
At Awendio Solaris, one of our planned research priorities is to use AI-guided experimentation to improve the stability of perovskite/silicon tandem cells. We aim to connect automated material deposition and characterisation with accelerated ageing tests, so that each candidate is assessed for efficiency, durability and reproducibility. Linking this work to pilot-line development would help us evaluate manufacturing feasibility and cost early. Scientists’ physical understanding and judgement remain essential throughout.
How much faster could we develop durable, affordable solar cells if every experiment whether successful or unsuccessful, helped us make better decisions about the next one?
Next: AI for PV manufacturing.
