ML-guided screening of chalcogenide perovskites as solar energy materials

This study presents a comprehensive, data-driven framework that integrates machine learning, a novel SISSO-derived tolerance factor, and sustainability metrics to screen and rank stable, experimentally feasible chalcogenide perovskites for next-generation solar energy applications.

The Gist

Imagine you are a master chef trying to invent a new, perfect cake. You know the basic recipe: flour, sugar, eggs, and butter. But you want to create a chocolate cake that is also vegan and gluten-free. The problem is, there are millions of possible ingredient combinations. If you tried baking every single one, you'd burn out your kitchen and run out of money before finding the winner.

This is exactly the problem scientists face when looking for new materials for solar panels. They are hunting for a specific type of crystal called a chalcogenide perovskite. Think of these as the "super-cakes" of the solar world—they could capture sunlight incredibly efficiently, last a long time, and be made from safe, earth-friendly ingredients. But finding the right recipe is incredibly hard because many ingredients that look like they should work, actually turn into a useless mess (like a brick) when you try to bake them.

This paper is about a team of scientists who built a super-smart, AI-powered "recipe filter" to find the best solar cake recipes without having to bake them all first.

Here is how their "Kitchen" works, step-by-step:

1. The New "Taste Test" (The SISSO Tolerance Factor)

Traditionally, chefs (scientists) used a simple ruler to measure if ingredients would fit together. They had a rule called the "Goldschmidt Tolerance Factor." It was like saying, "If the eggs are too big for the bowl, don't use them."

• The Problem: This old ruler was too blunt. It let in many ingredients that looked okay on paper but failed in the oven.
• The Fix: The team used an AI algorithm (called SISSO) to study thousands of past recipes. It invented a new, super-precise taste test (called $\tau^*$).
• The Analogy: Instead of just measuring the size of the egg, this new test checks the "personality" of the ingredients. It asks, "Do these specific atoms actually like each other enough to hold hands and form a stable structure?" This new test is much better at spotting the "bad eggs" before they even get into the mixing bowl.

2. The "Virtual Baker" (CrystaLLM)

Once the AI filtered out the bad ingredients, the team had a list of promising recipes. But just having the right ingredients doesn't guarantee the cake will rise.

• The Problem: Sometimes, even with good ingredients, the batter settles into a weird shape (a non-perovskite structure) instead of the perfect cube shape needed for solar power.
• The Fix: They used a generative AI model called CrystaLLM. Think of this as a "Virtual Baker" that has read millions of cookbooks. It looks at the list of ingredients and instantly "imagines" what the cake would look like if you baked it.
• The Result: The Virtual Baker said, "Hey, this combination might look like a brick, not a cake!" It filtered out the recipes that would form the wrong shape, leaving only the ones that would likely form the perfect solar-crystal structure.

3. The "Color Predictor" (Bandgap Estimation)

Now they have a list of cakes that might work. But is the cake the right color?

• The Concept: Solar panels need to absorb specific colors of light (like red or blue) to make electricity. If the material absorbs the wrong color, it's useless.
• The Fix: They used another AI (CrabNet) to predict the "color" (bandgap) of the material just by looking at the ingredients.
• The Analogy: It's like predicting that a cake with too much cocoa will be too dark (absorbing too much light) or one with too much sugar will be too pale (absorbing too little). They picked the "Goldilocks" recipes that absorb the perfect amount of sunlight.

4. The "Ethical & Economic Check" (Sustainability)

Finally, even if a cake tastes great and looks perfect, can you actually buy the ingredients?

• The Problem: What if the recipe requires "Unobtainium" or a rare element that only exists in a war-torn country? That's a bad recipe for the future.
• The Fix: They added a "Sustainability Score." They checked:
  • Supply Risk: Is this ingredient rare or hard to get? (Like trying to find truffles in a drought).
  • Toxicity: Is it safe? (No lead or radioactive uranium).
  • Cost: Is it affordable?
• The Result: They ranked the cakes not just by taste, but by how easy and safe it would be to make them for the whole world.

The Final Menu

After running all these filters, the team didn't just find one winner; they found a shortlist of the best candidates.

• The Star: They confirmed that BaZrS3 (a known material) is still a top contender, especially for high-end solar panels.
• The New Stars: They discovered several new, unexplored recipes (like CuHfS3 and EuYbSe3) that look incredibly promising. These are the "secret recipes" that no one has tried yet but the AI says might be the next big thing.

Why This Matters

Before this paper, finding these materials was like looking for a needle in a haystack by poking every piece of hay with a stick. It was slow, expensive, and frustrating.

This paper shows a new way: Use AI to look at the haystack, predict which pieces are needles, and only poke the most likely ones.

They aren't saying, "We have built the perfect solar panel." They are saying, "Here is a list of the top 30 recipes you should try baking in your lab first." This saves time, money, and energy, bringing us closer to a future where solar energy is cheaper, cleaner, and more efficient.

Technical Summary

1. Problem Statement

Chalcogenide perovskites (ABX₃ where X = S, Se, Te) are promising candidates for next-generation photovoltaics due to their intrinsic stability, tunable optoelectronic properties, and lack of toxic elements (unlike lead-halide perovskites). However, their experimental realization is severely limited by:

• Competing Phases: High thermodynamic stability of binary and polysulfide phases often prevents the formation of the desired perovskite structure.
• Structural Polymorphism: The existence of non-perovskite structures (e.g., edge-sharing octahedra) that are chemically similar but optoelectronically inferior.
• Synthetic Challenges: High synthesis temperatures and difficulties in scaling.
• Prediction-Reality Gap: Classical geometric descriptors (like the Goldschmidt tolerance factor) fail to reliably predict phase stability in chalcogenides due to their increased covalent character compared to oxides or halides. Furthermore, high-throughput Density Functional Theory (DFT) calculations are computationally expensive and do not always correlate with experimental synthesizability.

2. Methodology: A Multi-Stage ML Screening Pipeline

The authors propose a fully data-driven, experimentally grounded screening framework that integrates interpretable analytical descriptors, generative models, and machine learning (ML) to navigate the chemical space of hypothetical chalcogenide perovskites. The pipeline consists of four sequential stages:

A. Derivation of a New Tolerance Factor ($\tau^*$)

• Approach: Instead of relying on classical geometric factors, the authors used the SISSO (Sure Independence Screening and Sparsifying Operator) algorithm to derive a new analytical descriptor.
• Data: Trained on a curated dataset of experimentally reported halide and chalcogenide ABX₃ compounds (both perovskite and non-perovskite phases).
• Descriptor: The resulting formula, $\tau^*$, incorporates ionic radii ($r_A, r_B, r_X$) with higher-order terms and logarithmic size-mismatch terms to account for covalency and structural asymmetry.
• Criterion: Perovskite stability is defined as $\tau^* < 0.846$.

B. Generative Crystal Structure Prediction

• Tool: CrystaLLM, a large language model trained on over 3 million DFT-derived crystal structures.
• Function: Generates crystallographic information files (CIFs) for compositions passing the $\tau^*$ filter.
• Filtering: The generated structures are analyzed to ensure they adopt the specific corner-sharing $BX_6$ octahedral topology required for perovskites, filtering out edge-sharing motifs (common in $Pnma$ space groups) that do not constitute true perovskites.

C. Bandgap Estimation

• Tool: CrabNet, a composition-based deep learning model using attention mechanisms.
• Training: Trained on a diverse dataset of hybrid halide perovskites, chalcogenide perovskites, and chalcogenide semiconductors to ensure generalizability.
• Output: Predicts the optical bandgap ($E_g$) for the 54 structurally validated candidates.

D. Experimental Plausibility and Sustainability Assessment

• Synthesizability: Uses a pre-trained Graph Convolutional Neural Network (GCNN) to calculate a Crystal-Likeness Score (CLS). This PU (Positive-Unlabeled) learning approach predicts the likelihood of a structure being experimentally realizable based on known synthesized crystals.
• Sustainability: Calculates Supply Risk (SR) using the Herfindahl-Hirschman Index (HHI) for market concentration and Environmental, Social, and Governance (ESG) scores from World Bank data.
• Multi-Objective Ranking: Candidates are ranked based on their suitability for single-junction ($E_{opt} \approx 1.34$ eV) and tandem ($E_{opt} \approx 1.71$ eV) photovoltaic architectures, balancing bandgap deviation, supply risk, and synthesizability.

3. Key Contributions

• Novel Descriptor ($\tau^*$): Introduced a SISSO-derived tolerance factor that significantly outperforms established metrics (Goldschmidt, Jess, Bartel) in predicting chalcogenide perovskite stability, achieving 91.2% accuracy and 95.0% precision on experimental data.
• Hybrid Screening Strategy: Demonstrated a transferable workflow that combines geometric stability, generative structural prediction, and experimental plausibility scoring, reducing reliance on expensive DFT calculations.
• Sustainability Integration: Pioneered the inclusion of socio-economic supply risk metrics early in the materials discovery pipeline to identify not just high-performance, but also scalable and sustainable materials.
• Identification of New Candidates: Identified several previously unexplored chalcogenide perovskites with high potential for photovoltaic applications.

4. Key Results

• Performance of $\tau^*$: The new descriptor reduced false positives significantly compared to $t_{Jess}$ and $t_{Bartel}$ while maintaining high recall. It correctly identified the stability window for 181 compositions out of 1,392 possible candidates.
• Structural Refinement: After applying CrystaLLM to verify the corner-sharing topology, the candidate pool was reduced to 54 compounds.
• Bandgap Analysis: The 54 candidates exhibited a mean bandgap of 1.86 eV, making them ideal for top cells in tandem configurations. The model showed a Mean Absolute Error (MAE) of 248 meV on the test set.
• Top Candidates:
  • BaZrS₃ remains the benchmark, showing optimal properties for tandem cells.
  • New Promising Materials: CuHfS₃, CuZrSe₃, EuYbSe₃, EuScS₃, and LaTbS₃ were identified as high-priority targets.
  • Synthesizability: Most top candidates had a Crystal-Likeness Score (CLS) > 0.88, indicating high statistical similarity to known synthesized materials.
• Trade-offs: While rare-earth-based compounds (e.g., Eu, Tb) showed excellent geometric and electronic properties, they often carried higher supply risks. Copper-based compounds (e.g., CuHfS₃) offered a better balance of low supply risk and suitable bandgaps.

5. Significance

This work addresses the critical bottleneck in materials discovery where theoretical predictions often fail to translate to experimental reality. By integrating interpretable physics-based descriptors with generative AI and socio-economic constraints, the authors provide a robust framework for discovering materials in "chemically constrained" spaces.

The study highlights that:

1. Geometric stability is necessary but insufficient: Topological constraints (corner-sharing) and experimental plausibility (CLS) are critical filters.
2. Sustainability is a design parameter: Supply chain risks must be evaluated alongside optoelectronic performance to ensure viable large-scale deployment.
3. Data-driven workflows can bridge the gap: This pipeline offers a prioritized list of targets for experimentalists, significantly reducing the time and cost associated with trial-and-error synthesis in the search for stable, non-toxic solar absorbers.

The authors emphasize that while these are prioritized targets, experimental verification is still required, particularly regarding oxidation state stability (e.g., Cu²⁺ vs Cu⁺) and redox thermodynamics, which are not explicitly modeled in the current workflow.