Everyone’s scrambling for better solar energy, tech that’s efficient, stable, and cheap. Perovskites look like a great candidate. But the hype around using deep learning to find new perovskite materials is getting out of hand, and a lot of what you hear is just plain wrong.

Myth 1: Deep Learning Can Magically Design New Perovskites from Scratch

The biggest myth is that some AI can just sit there and invent completely new perovskite structures without any human guidance or knowledge of basic chemistry. It’s a nice sci-fi image, but that’s not how AI works in science today. At its core, deep learning is a pattern-finding machine, and a very good one at that.

Its power comes from finding non-obvious connections in huge amounts of data. For perovskites, that means feeding a model tons of existing experimental and computational data, what they’re made of, how they were made, and how they performed. For instance, you can train a model on thousands of known ABX3 perovskites, linking the elements at the A, B, and X sites to the material’s bandgap. The model learns those patterns and can then predict the properties of untested combinations of those same elements. It’s making an educated guess based on what it’s already seen, just interpolating and extrapolating within a known chemical space.

Look at the work presented at the 2025 Materials Research Society (MRS) Fall Meeting. Researchers from NREL showed off a deep learning model that could predict the thermodynamic stability of halide perovskites with over 92% accuracy. But it wasn’t pulling ideas out of thin air. The model worked by analyzing known chemical descriptors and crystal structures from density functional theory (DFT) calculations and experimental databases. That foundational data is what keeps the model’s predictions grounded in reality. Without it, the AI has no rules to follow. It can’t invent quantum mechanics.

So, think of deep learning as an accelerator, a tool for rapidly sorting through thousands of possibilities to flag the handful that are actually worth a chemist’s time in the lab. The human researcher, who actually understands the physics, is still the one who has to frame the problem, judge the AI’s output, and design the real-world experiments.

Myth 2: You Need Petabytes of Data for Effective Deep Learning in Perovskite Research

People in deep learning love to say “more data is better.” This has led to the mistaken belief that you need gigantic, petabyte-scale datasets to do anything useful in a specialized field like materials discovery. The truth is a lot more complicated, especially when each data point comes from a difficult, expensive lab experiment.

Most labs working on new perovskites have small, hard-won datasets, not petabytes of perfectly organized sample information. The good news is that many of the most successful deep learning applications in this field use these smaller, high-quality datasets. They just get clever with techniques like data augmentation or transfer learning. For example, a 2023 paper in npj Computational Materials showed you could effectively predict perovskite stability with just a few hundred data points by enriching them with features derived from computation. They didn’t need millions of samples. They just needed the right features.

And “data” doesn’t just mean physical lab results. We can use high-throughput simulations like DFT to generate huge amounts of synthetic data about a material’s properties. While this data isn’t a direct experimental measurement, you can use it to pre-train a model, giving it a basic ‘education’ in material physics before showing it your smaller, more precious experimental dataset. This lets you explore a much bigger chemical space than you could with experiments alone.

Active learning is also a huge part of the solution. Instead of just gathering data randomly, an active learning algorithm intelligently picks the next experiment that will be most informative for the model. It focuses your lab work where it will have the most impact. The Materials Efficiency Group at the University of Cambridge has used these methods to discover new inorganic materials an order of magnitude faster than before. That’s a long way from needing petabytes.

Myth 3: Deep Learning Models Are Black Boxes with No Explanations

For a long time, scientists were skeptical of deep learning because of the “black box” problem. The models were accurate, but you had no idea *how* they reached their conclusions. For a materials scientist who needs to understand the physics, a prediction without an explanation is not very useful. But the idea that deep learning can’t explain itself is becoming a myth.

The field of Explainable AI (XAI) has produced tools that let us look inside the models. For perovskite research, this means we can finally ask *why* a model predicted a certain bandgap or stability. Techniques like SHAP and LIME can show us exactly how much each input feature contributed to a prediction. So if a model flags a new perovskite composition as highly stable, XAI can tell you if that’s because of the A-site cation’s ionic radius, the electronegativity of the B and X sites, or something else. The model stops being just a predictor and starts becoming a tool for generating new scientific hypotheses.

Some models are also just easier to interpret by design. Graph Neural Networks (GNNs) which are a natural fit for describing crystal structures, can learn physically meaningful representations of atoms and bonds. By looking at these learned representations, you can get real insights. A 2024 paper in The Journal of Physical Chemistry Letters used a GNN that not only predicted the formation energy of hybrid perovskites but also let the researchers see how specific rotations of the organic cation were linked to better stability. That’s not a black box, it’s a direct suggestion for what to try next in the lab.

We’re not at the point where we can get a perfect physical explanation for every prediction from every model. But the idea that deep learning only gives answers without understanding is flat-out wrong. The community is working hard to make sure these tools provide scientific insight, not just numbers.

Myth 4: Deep Learning Will Replace Experimental Chemists and Materials Scientists

The fear that AI will make human jobs obsolete pops up everywhere, and materials science is no different. You hear people worry that since AI can screen materials so fast, we won’t need experimental chemists working on perovskites anymore. This completely misunderstands what an AI is and what a scientist does.

Deep learning is a tool that makes scientists better, it doesn’t replace them. Who do you think creates the high-quality data the models train on? Who actually goes into the lab to synthesize the new composition the AI suggested? Who runs the XRD and SEM to see if the prediction was right? Who has to troubleshoot the hundred little things that go wrong during a synthesis? (And trust me, things always go wrong). It’s a human scientist. The hands-on skill, intuition, and critical thinking required for lab work just can’t be replicated by today’s AI.

The whole scientific method depends on human expertise. You need a person to form a hypothesis, design a smart experiment, collect the data, and then interpret what it all means. AI can speed up parts of that cycle, especially the analysis and prediction steps. It helps scientists see the bigger picture and make smarter bets on what to try next. But the Materials Genome Initiative, the big US government project in this space, has always been about integrating computation and experiments, not having one replace the other. Their 2026 strategic plan is all about training a workforce that can use both. In my view, the scientists who learn to use these tools will be the ones making the big discoveries, and the ones who don’t are the ones who should be worried.

Myth 5: All Deep Learning Models for Perovskites Are Equally Reliable

The hype around AI can make people accept any result that comes from a “deep learning model” as gospel. This is a huge mistake. The belief that any model built for perovskites is automatically trustworthy is a dangerous oversimplification that will lead to a lot of wasted time and money.

A model’s reliability is tied directly to a few key things. First and most important is the **training data**. A model trained only on lead-halide perovskites will give you garbage predictions for lead-free double perovskites because it has no idea what those look like. If your data is biased toward materials made a certain way, your model will be biased too. Second, the **model architecture** has to fit the problem. A simple neural network might work for a simple prediction, but you’ll need something more complex like a GNN to understand the details of a crystal structure. Third, you need rigorous **validation and uncertainty estimates**. A model that’s “95% accurate” isn’t very helpful if you don’t know the uncertainty on any single prediction. Is that new material a sure thing or a long shot? Researchers at MIT have been very clear about this: a prediction without an error bar is scientifically useless.

And what’s the “best” model for predicting a bandgap might be a terrible model for predicting long-term stability. Every new problem requires careful tuning, feature selection, and testing against the right benchmarks. You can’t just download a pre-trained model and expect it to work on your specific perovskite problem. The field is moving towards more open-source code and reproducible methods for exactly this reason, so we can all kick the tires on these models and see which ones actually hold up.

Deep learning has real potential to change how we discover perovskite materials and speed up the arrival of next-gen solar tech. But for that to happen, we have to use these tools with a clear-eyed view of what they can and can’t do, and ignore the myths.