Supervised Machine Learning for Predicting Power Conversion Efficiency in Perovskite Solar Cells: A Systematic Review and Quantitative Meta-Analysis

Imagine you are trying to bake the world's most perfect chocolate chip cookie. But instead of just flour and sugar, your recipe involves thousands of secret ingredients, a mysterious oven that changes temperature every second, and a mixing bowl that reshapes itself while you stir. This is what scientists face when building Perovskite Solar Cells (PSCs). These are the shiny, next-generation solar panels that could power our future, but figuring out the perfect recipe to make them super efficient is a nightmare of trial and error.

Enter Supervised Machine Learning (ML). Think of this not as a magic crystal ball, but as a super-smart, tireless apprentice chef who has read every single recipe book ever written. Instead of baking a thousand cookies to see which one tastes best, this apprentice looks at the data from previous bakes and predicts exactly which mix of ingredients will create the perfect cookie.

The Great Cookie Contest: Who Wins?

The authors of this paper decided to act like food critics. They didn't just taste one cookie; they gathered 36 different studies (published between 2015 and 2025) where scientists used these AI chefs to predict the Power Conversion Efficiency (PCE)—basically, how much sunlight the solar cell turns into electricity.

They put all these AI models into a giant arena to see which one actually works best. Here is the scoreboard they found:

• The Tree-Based Champions (Random Forest & XGBoost): These models are like a team of expert chefs who vote on the recipe. They look at the data and say, "If the temperature is high and the mixing speed is low, we get a great result." The paper found that these tree-based ensemble methods are the current kings of the kitchen. When tested on standard data tables, they consistently hit an R² value between 0.80 and 0.90. In plain English, this means they are incredibly accurate, getting the prediction right about 80% to 90% of the time.
• The Neural Network Contenders (Artificial Neural Networks): These are like deep-learning robots that try to mimic the human brain. The paper suggests they are getting stronger, but they have a catch: they need a massive amount of data to shine. If you have a small dataset (fewer than 300 samples), these robots tend to stumble. They only start to compete with the tree-based champions when you feed them large datasets (over 1,000 samples).
• The Underperformers: The paper explicitly rules out some other methods as the "go-to" solution for this specific job. Simple Linear Models (like basic math equations) and Support Vector Regression were found to be much weaker, often failing to capture the complex, twisty relationships in the solar cell data. Their accuracy was significantly lower, with R² values often dropping below 0.60.

The "Goldilocks" Rule for Data Size

One of the coolest discoveries in this paper is how the size of your data changes which AI chef you should hire. The authors mapped this out like a guide for choosing the right tool:

1. Small Datasets (<300 samples): If you only have a few recipes to work with, Random Forest is your best bet. It's stable and doesn't get confused easily.
2. Medium Datasets (300–800 samples): This is the sweet spot for XGBoost and other Gradient Boosting models. They need a bit more data to tune their knobs perfectly, but once they do, they perform amazingly well.
3. Large Datasets (>1,000 samples): Now you can bring out the heavy hitters: Neural Networks. With enough data to learn from, they can start to outperform the others by spotting hidden patterns that simpler models miss.

The Problem with the "Kitchen"

Even though these AI chefs are talented, the paper points out that the kitchen is a bit messy.

• No Standard Recipes: Different scientists use different ways to measure their results. Some use "R," others use "R²," and some don't even tell you how they tested their models. It's like one chef measuring cups in milliliters and another in ounces. This makes it hard to compare who is truly the best.
• The "Train-Test" Trap: About 52% of the studies just split their data in half (one half to learn, one half to test). The paper argues this isn't strict enough. A better way is k-fold cross-validation (used in about 41% of studies), where the model is tested on many different slices of data to ensure it's not just memorizing the answers.
• Black Boxes: Many models are "black boxes," meaning they give an answer but don't explain why. The paper suggests we need to move toward Explainable AI (using tools like SHAP) so scientists can understand which ingredient (like annealing temperature or material composition) actually matters most.

The Future Roadmap

The authors aren't just saying "AI is good." They are drawing a map for the future. They suggest a five-step plan to take solar cell research from "guessing" to "autonomous discovery":

1. Clean Up the Data: Create open, standard databases so everyone speaks the same language.
2. Smart Features: Don't just throw random numbers at the AI; give it features that make physical sense (like bandgap energy).
3. Rigorous Testing: Use strict validation methods so we know the models actually work, not just on paper.
4. Physics-Informed AI: This is a big one. Instead of letting the AI guess blindly, we should feed it the actual laws of physics (like how electricity moves). This ensures the AI doesn't suggest a solar cell that looks great on a computer but is physically impossible to build.
5. Self-Driving Labs: Imagine a robot that uses the AI to design a solar cell, builds it, tests it, and then feeds the results back to the AI to try again. This "closed-loop" system could speed up discovery massively.

The Bottom Line

This paper doesn't claim that AI has "solved" solar cells. Instead, it suggests that tree-based models like Random Forest and XGBoost are currently the most reliable tools for predicting solar cell efficiency, especially when working with the medium-sized datasets scientists usually have. It warns us that we need better data standards and more transparent testing to trust these results fully. But if we follow their roadmap—combining smart AI with the laws of physics—we might just see the perfect solar cell recipe much sooner than we thought.