Machine Learning Unlocks Atomic Mechanisms of Passivation in Perovskite Solar Cells ✛ ✎ ✦
Context
Hybrid halide perovskites are cutting-edge materials for solar cells, but their degradation under moisture, heat, and light limits their use. Molecular surface passivation with aminosilanes boosts stability, yet the atomic-scale mechanisms remained murky. Machine learning via interatomic potentials opens a path to model such complex systems, but it must overcome challenges in transferring knowledge between disparate chemical environments.
Methods
The researchers employed the MACE architecture and a continuous fine-tuning protocol similar to that used for large language models. First, based on numerical simulations and quantum calculations rooted in the Schrödinger equation, they compiled the extensive 'hyP-26' dataset of diverse perovskite structures containing silicon, carbon, oxygen, and hydrogen. Then, precision tuning of the potential was performed on a specialized dataset of AEAPTMS molecule coordination on the surface. They compared different starting models: one pre-trained mainly on inorganic crystals, and another with broader coverage of organic and surface systems.
Results
For the narrowly pre-trained model, intermediate fine-tuning on 'hyP-26' significantly improved force and adsorption energy prediction accuracy, slashing the error by more than half (from 240 to 90 meV/Å). For the more universal model, direct transfer yielded comparable results. Analysis showed that apparent accuracy without the intermediate step often arises from error cancellation. Molecular dynamics numerical experiments with varying passivator coverage revealed that above 0.8 molecules/nm², the organosilicon molecules switch from a multidentate to a vertical orientation, burrow into the surface, and disrupt octahedral order, leading to increased octahedral tilt angles and elongated Pb–X bonds.
Implications
The work provides a blueprint for machine learning in materials science: combining general and specialized datasets enables more accurate descriptions of heterogeneous systems without training from scratch. The atomistic insight explains why excessive passivator application degrades film crystallinity and highlights the importance of steric effects in designing new molecules, including the balance of carbon chains and functional groups.
Future development
In the future, such potentials will enable virtual screening of thousands of passivating agents even before lab tests. Integration with generative models will speed up the creation of optimized molecules, and accounting for oxygen-containing groups and hydrogen bonds will broaden the horizons of computational materials design for photovoltaics.
Impact
The methodology will influence the development of stable perovskite solar cells and LEDs, as well as any field requiring modeling of organic molecule interactions with inorganic surfaces, including catalysis and sensors where silicon substrates are critical.
Next steps
Immediate next steps include accounting for chemical reactions between perovskite cations and the passivator, and examining the impact of various surface defect types, which will demand even more detailed numerical calculations.
Key open problems
The research is directly tied to predicting the long-term stability of functional materials under real-world conditions — a key challenge at the intersection of quantum chemistry and statistical physics. The success of fine-tuning protocols also touches on a fundamental question about knowledge transfer in neural network models, echoing the catastrophic forgetting problem.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study presents a multi-step training protocol for interatomic potentials to model surface passivation of halide perovskites. Using the hyP-26 dataset and specialized coordination data for AEAPTMS, the potential was fine-tuned, revealing how molecular orientation and surface structural changes depend on coverage.
Where its strength is
The strength of the work lies in its systematic approach to transferring knowledge between disparate datasets, demonstrating the importance of intermediate fine-tuning to mitigate error compensation and achieve reliable predictions for complex heterogeneous systems.
Where it could go next
- Transferring the continuous fine-tuning protocol to other material classes, such as solid electrolytes for batteries where accurate interface description is also crucial, could be tested, drawing on the work's insight that data quality matters more than quantity.based on: Solid-State Batteries: The Secret Isn't in Data Volume arXiv:2603.07425
- Extending the approach to discover optimal passivating molecules by integrating active learning or Bayesian optimization, akin to searching for recipes in chemical space, could accelerate the design of new coatings.based on: A Chemist Computer Hunts for the Recipe for a Better Solar Cell arXiv:2603.02717
- Using ideas from model adaptation to new chemical environments with minimal additional data, as in the work on tuning quantum calculations for caloric materials, could reduce the cost of creating specialized potentials for other surfaces.based on: The neural network tuner: how five quantum calculations open the door to Freon-free fridges arXiv:2606.25742
Worth trying
- Test how much the protocol with intermediate fine-tuning improves transfer to surfaces with other terminal groups (e.g., methylammonium or formamidinium), using only a few dozen new calculations.
- Apply error compensation analysis to assess prediction reliability when extrapolating to higher coverages than those in the training set.
- Combine the resulting potential with generative molecular design methods to automatically propose candidates with an optimal balance of steric and binding properties.
Why this may matter
Advances in accurate and cost-effective modeling of perovskite interfaces could accelerate the development of stable solar cells, and the proposed training protocol may serve as a general template for adapting universal interatomic potentials to specialized materials science tasks.
Nearby works in our archive
- A Chemist Computer Hunts for the Recipe for a Better Solar Cellsimilarity 0.68
- How Clever Sampling Teaches Metal to Predict Strengthsimilarity 0.63
- How a Neural Network Learns from Quantum Chemistry's Mistakessimilarity 0.62
- Teaching Neural Networks to Read Molecules at a Glancesimilarity 0.62
- Neural Network Sculptor: Reading a Gravitational Lens from Four Pointssimilarity 0.60
- Solid-State Batteries: The Secret Isn't in Data Volumesimilarity 0.60
- The neural network tuner: how five quantum calculations open the door to Freon-free fridgessimilarity 0.59
The closest work in the archive sits at 0.68 by meaning — the usual level for our corpus.
The name 'perovskite' comes from Russian mineralogist Lev Perovski, and modern hybrid perovskites can contain up to six different elements, enabling fine-tuning of properties. Simulating a single molecule on the surface required calculations involving over 3,000 atoms — comparable to the size of an entire nanocrystal.
Key numbers
- RMS force error before fine-tuning: 240 meV/Å
- RMS force error after fine-tuning: 90 meV/Å
- Surface passivator concentration: 0.03–1.28 molecules/nm²
- Fraction of defective octahedra at maximum coverage: 7.1%
- Molecular dynamics temperature: 450 K