Charting the thermodynamic stability of hybrid perovskite alloys with machine learning

This study employs a two-level machine learning strategy combining graph neural network potentials and direct energy predictors to map the thermodynamic stability of (Cs/FA)Pb(Br/I)₃ and (Cs/FA)Sn(Br/I)₃ perovskite alloys, revealing that Sn-based systems have narrower stable composition regions than Pb-based ones and that maximum stability occurs at high iodine content.

The Gist

Imagine you are a master chef trying to create the perfect, most stable cake. You have four main ingredients: Cesium (Cs), Formamidinium (FA), Lead (Pb) or Tin (Sn), and a mix of Bromine (Br) and Iodine (I). By mixing these in different amounts, you can create a "hybrid perovskite" cake that turns sunlight into electricity.

The problem is that there are billions of possible recipes. If you tried to bake and test every single one in a real kitchen (or with real computers using standard methods), it would take longer than the age of the universe. Some recipes taste great but fall apart quickly (unstable), while others are rock-solid but don't taste good (low efficiency).

This paper is about a team of scientists who built a super-smart AI kitchen assistant to figure out which recipes are the best without actually baking billions of cakes.

Here is how they did it, using simple analogies:

1. The Two-Step AI Strategy

The scientists realized that testing every recipe was too slow, so they built a two-level system:

• Level 1: The "Smart Taster" (MACE Model)
  Think of this as a highly trained chef who can look at a raw, uncooked batter and predict exactly how it will settle and taste once baked. This AI was trained on a small number of real, expensive computer experiments (called DFT). It learned the rules of physics so well that it can "relax" (settle) a structure almost instantly, saving about 1,000,000 times the time it would take a standard computer.
• Level 2: The "Crystal Ball" (Direct Relaxation Model)
  Even the "Smart Taster" is too slow if you need to check billions of recipes. So, the scientists built a second, even faster AI. This one looks at the raw, uncooked batter and predicts exactly what the "Smart Taster" would have said after baking it. It skips the baking step entirely. This second step saved another 1,000 times in speed.

The Result: Together, these two AIs allowed the team to taste-test 2 billion different recipes in a fraction of the time it would have taken before.

2. The Map of Stability

Using this super-fast system, they created a "stability map" for two types of cakes:

1. Lead-based cakes: (The current champions of efficiency).
2. Tin-based cakes: (The eco-friendly, non-toxic alternative).

They looked at the map to see which recipes were "stable" (won't fall apart) and which were "unstable" (will crumble or separate).

3. The Big Discoveries

• The Lead vs. Tin Showdown: The map revealed that the Lead-based cakes have a wide, safe zone where you can mix ingredients and still get a stable result. However, the Tin-based cakes are much more fragile. Their "safe zone" is very narrow. If you try to mix them too much, they tend to fall apart. This explains why making non-toxic solar cells is so hard; you have very few options for tweaking the recipe without breaking it.
• The "Middle" is Unstable: You might think mixing everything in the middle (50% of this, 50% of that) would be the most stable, like a perfect balance. The map showed the opposite. The most stable spots are usually at the edges (high Iodine content), while the center of the map is a "danger zone" where the material wants to separate into different parts.
• Heat Helps (A Little): They checked the map at room temperature and at high baking temperatures (150°C). While heat did make the stable zones slightly bigger, the fundamental problem with Tin-based cakes (their narrow safe zone) remained.

4. Why This Matters

The paper doesn't claim to have invented a new solar cell today. Instead, it provides a roadmap.

• For scientists trying to make solar cells, it says: "Don't waste time trying to mix Tin-based cakes in the middle of the recipe book; they won't work. Stick to the edges where the map says it's safe."
• It confirms that while we want to use Tin to avoid toxic Lead, the laws of physics make Tin-based alloys much harder to stabilize than Lead-based ones.

In short: The scientists built a super-fast AI that mapped out the "safe zones" for solar cell ingredients. They found that while Lead-based mixtures are flexible and stable, the eco-friendly Tin-based mixtures are much pickier and harder to keep together, guiding future researchers on where to look for the next breakthrough.

Technical Summary

Technical Summary: Charting the thermodynamic stability of hybrid perovskite alloys with machine learning

Problem Statement
Hybrid perovskite solar cells (PSCs) have achieved power conversion efficiencies rivaling silicon, yet their commercialization is hindered by long-term operational stability issues and lead toxicity concerns. Compositional engineering using quaternary alloys, specifically $(Cs/FA)Pb(Br/I)_3$ and the lead-free $(Cs/FA)Sn(Br/I)_3$ (where FA is formamidinium), offers a pathway to tune properties and improve stability. However, the vast configurational space of these multicomponent systems, combined with the complexity of orientable organic cations, has impeded accurate thermodynamic modeling. Traditional Density Functional Theory (DFT) is computationally prohibitive for sampling the necessary configurations to map free energy landscapes, while existing approximations like cluster expansions fail to capture molecular cation orientations, and Special Quasirandom Structures (SQS) introduce biases due to limited configuration sampling.

Methodology
The authors present a machine learning (ML) accelerated atomistic modeling workflow designed to compute free energy landscapes across the full composition space of these quaternary alloys. The approach employs a two-level ML strategy to overcome computational bottlenecks:

1. MACE Interatomic Potentials: The authors trained MACE (Higher-order Equivariant Message Passing Neural Networks) potentials on DFT data. These potentials were used to perform efficient structure relaxations, accelerating computations by approximately six orders of magnitude compared to DFT. The training data included $2 \times 2 \times 2$ supercells from four distinct phases ($Pm\bar{3}m$, $P4/mbm$, $I4/mcm$, and $Pnma$) with randomized cation/anion distributions and FA molecular orientations.
2. Direct Relaxation Models (KRR): To further accelerate the sampling required for free energy calculations, secondary Kernel Ridge Regression (KRR) models were trained on the MACE-relaxed data. These models predict relaxed energies directly from unrelaxed atomic structures, bypassing the need for explicit relaxation during sampling. This step provided an additional three orders of magnitude speedup.
3. Active Learning: To ensure accuracy, particularly in the low-energy regime critical for stability analysis, an active learning scheme was employed. This involved iterative energy minimization Monte Carlo (MC) simulations to identify and correct predictions for low-energy alloy configurations, reducing mean absolute errors (MAEs) significantly.
4. Free Energy Calculation: The Wang-Landau algorithm was utilized to sample the density of states and compute Helmholtz free energies of mixing at 300 K (and 150 °C for synthesis conditions). This allowed for the construction of convex hulls and curvature analysis to assess thermodynamic stability against phase separation.

Key Contributions

• Workflow Development: The paper establishes a rigorous multi-stage ML workflow capable of handling the complexity of quaternary hybrid perovskites, including the treatment of orientable organic cations, which previous methods struggled to model effectively.
• Speedup: The combined approach achieves a total speedup of approximately nine orders of magnitude compared to traditional DFT, enabling the evaluation of over two billion energy configurations necessary for comprehensive free energy mapping.
• Comparative Analysis: The study provides the first comprehensive thermodynamic stability comparison between lead-based $(Cs/FA)Pb(Br/I)_3$ and lead-free $(Cs/FA)Sn(Br/I)_3$ quaternary alloys across multiple phases.

Results

• Model Accuracy: The MACE potentials achieved MAEs of 2.81 meV/f.u. and 3.07 meV/f.u. for Pb- and Sn-based systems, respectively, compared to single-point DFT calculations on relaxed structures. The final direct relaxation models (after active learning) achieved MAEs of approximately 3.2 meV/f.u. on minimum-energy configurations.
• Stability Landscapes:
  • Pb-based System: Exhibits broader stable composition regions. In the cubic ($Pm\bar{3}m$) phase, a low-energy region exists in the center of the composition space. Stable regions on the convex hull extend across various A-site substitutions and up to 60% Br concentration depending on the Cs/FA ratio.
  • Sn-based System: Shows narrower stable composition regions compared to its Pb counterpart. The central low-energy well observed in the Pb-based cubic phase is absent; the lowest energies are found at the Cs-free edge. The convex hull indicates fewer options for compositional engineering, with stability largely confined to high Iodine content.
  • Temperature Effects: Stability regions expand at elevated temperatures (150 °C), consistent with entropic stabilization, but the fundamental disparity between Pb and Sn systems remains.
• Entropy and Stability: The analysis reveals that entropy-driven stabilization is limited. Maximum stability occurs at high Iodine content, and no significant stabilization is observed near the center of the composition space where mixing entropy is highest. This suggests that mixing alone cannot stabilize the central quaternary compositions against phase separation.
• Experimental Validation: The results for the Pb-based system show qualitative agreement with high-throughput experimental data from Wang et al., particularly regarding the stability of low-Br, low-Cs regions in the cubic phase. Discrepancies in the orthorhombic phase are attributed to non-equilibrium synthesis conditions and potential phase identification challenges between similar crystal structures.

Significance and Claims
The paper claims that its results guide the design of stable perovskites by mapping the fundamental thermodynamic limits of these alloys. The authors conclude that the reduced thermodynamic stability of tin-based hybrid perovskites arises from fundamental differences in their free energy landscapes compared to lead-based counterparts, specifically the absence of central low-energy regions and narrower stable composition ranges. Consequently, the study suggests that future experimental efforts to optimize $(Cs/FA)Sn(Br/I)_3$ should focus on compositions already identified as favorable for the Pb-based system, rather than exploring the central composition space where stability is not thermodynamically favored. The work highlights that while ML can accelerate discovery, the intrinsic thermodynamic instability of certain quaternary regions cannot be overcome solely by mixing entropy.