🔬 materials science

Machine Learning-Accelerated Band-Edge Engineering of Pnictogen Chalcohalide Solid Solutions for Solar Energy Technologies

This study employs machine learning combined with first-principles calculations to map the tunable band-edge positions of pnictogen chalcohalide solid solutions across their full compositional range and surface terminations, identifying specific formulations suitable for various solar energy applications and revealing that facet selection is a critical design parameter comparable to chemical substitution.

Original authors: Cibrán López, David Rovira, Edgardo Saucedo, Claudio Cazorla

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cibrán López, David Rovira, Edgardo Saucedo, Claudio Cazorla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The quest for better solar energy and cleaner fuels often leads scientists to the atomic scale, searching for materials that can catch sunlight and turn it into electricity or chemical energy. At the heart of this search are semiconductors, a special class of solids that act as the bridge between light and electrical current. For a material to be useful in these roles, its internal energy levels must be just right: they need to be positioned so that when light hits them, the resulting electrical charges can flow out to do work, or push chemical reactions forward. If these energy levels are even slightly off, the material might absorb light but fail to release the energy, or it might simply let the charges cancel each other out before they can be used. Finding the perfect material is like tuning a radio to a specific station; the signal must be clear and strong, and the frequency must match exactly what the device needs to receive.

In recent years, a family of materials known as pnictogen chalcohalides has emerged as a promising candidate for these tasks. These compounds are made from elements that are common in the Earth's crust, are non-toxic, and can be processed at relatively low temperatures, making them attractive for large-scale energy applications. They naturally possess a wide range of energy gaps, the distance between their lowest and highest energy states, which can be adjusted by changing their chemical makeup. However, these materials are not just simple blocks of atoms; they have distinct surfaces, much like the different faces of a crystal. The behavior of electrons on one surface can differ significantly from another, even if the material inside is the same. This creates a massive puzzle for researchers: with so many possible combinations of ingredients and so many different ways the surface can be arranged, it is impossible to test every single possibility in a laboratory. The number of potential recipes is simply too vast to explore one by one.

To solve this problem, a team of researchers combined the precision of computer simulations with the speed of machine learning. They focused on a specific group of these materials, which can be thought of as a solid mixture where the ingredients can be swapped in and out in any proportion. The team used powerful computers to calculate the exact energy levels for a large number of these mixtures on two of their most stable surfaces. Because calculating every single possible combination would take too long, they trained a computer program to recognize patterns in the data they did calculate. This artificial intelligence model learned to predict the energy levels for thousands of new, untested mixtures with high accuracy, effectively mapping out the entire landscape of possibilities without needing to build them all.

The results of this digital exploration revealed a powerful new way to control these materials. The researchers found that by simply changing the ratio of ingredients, they could shift the energy levels by more than one electronvolt, a significant amount that allows for precise tuning. More surprisingly, they discovered that the surface itself acts as a major control knob. Even for the exact same mixture of ingredients, the energy levels on one surface could be shifted by up to 0.6 electronvolts compared to the other surface. This means that choosing which face of the crystal is exposed is just as important for designing a device as choosing the chemical recipe. The two surfaces, while nearly identical in their stability, behave very differently when it comes to interacting with light and electricity.

When the researchers applied these findings to real-world applications, the differences became clear. For the task of splitting water to produce hydrogen fuel, the surface orientation proved critical. One surface orientation was found to be capable of driving the necessary chemical reactions to split water into hydrogen and oxygen, provided the right mix of ingredients was used. The other surface orientation, despite being made of the same material, failed to straddle the energy requirements needed for this reaction, rendering it inactive for this specific purpose. The study identified specific combinations of elements that could drive not only hydrogen production but also the creation of ammonia, methane, and hydrogen peroxide, effectively mapping out a menu of fuel-making options based on the material's composition and surface.

In the realm of solar cells, the findings offered a cautionary but useful insight. The researchers tested how well these materials would work with common contact layers used in solar panels to collect electricity. They found that while the materials could work well with certain layers to collect positive charges, they did not align properly to collect negative charges with the same standard materials. This suggests that while these new materials are excellent at absorbing light and generating charges, building a complete solar cell will require finding or designing a new type of contact layer specifically for collecting the negative charges. The study concludes that these materials offer a highly flexible platform for energy technologies, but their success depends on carefully engineering both the chemical mix and the specific surface exposed to the environment. By using machine learning to navigate this complex space, the researchers have provided a clear roadmap for how to harness these materials for a cleaner energy future.

Technical Summary: Machine Learning-Accelerated Band-Edge Engineering of Pnictogen Chalcohalide Solid Solutions

Problem Statement
Pnictogen chalcohalides (MChX; M=Bi,Sb; Ch=S,Se; X=I,Br) are emerging semiconductors with earth-abundant constituents, robust thermodynamic stability, and optimal band gaps (1.2–2.1 eV) for solar energy applications. While ternary MChX compounds show promise, their solid solutions ( $Bi_xSb_{1-x}S_ySe_{1-y}I_zBr_{1-z}$ ) offer a vast compositional space for continuous property tuning. However, the sheer number of possible atomic configurations and surface terminations makes a systematic assessment of band-edge positions via exhaustive first-principles density functional theory (DFT) computationally prohibitive. Furthermore, experimental synthesis of all potential compositions is impractical. A critical gap exists in understanding how band-edge positions vary with stoichiometry and surface termination, which is essential for optimizing these materials in photovoltaics (PV) and photocatalysis.

Methodology
To address the computational bottleneck, the authors developed a machine learning (ML)-accelerated framework combining DFT with predictive modeling:

Dataset Generation: A database of 125 ternary MChX compounds was constructed using DFT. Geometry relaxations were performed using the PBEsol functional, while band alignments were calculated using the range-separated hybrid functional HSEsol with spin-orbit coupling (SOC) corrections.
Modeling Approach: Due to the prohibitive cost of HSEsol+SOC calculations on large supercells required for random solid solutions, the study employed the Virtual Crystal Approximation (VCA). This method assumes virtual atoms on disordered sites, interpolating electronic properties between isoelectronic components (Bi/Sb, S/Se, I/Br), which is valid given their similar bonding characters.
Machine Learning: A multilayer perceptron (MLP) model was trained on the 125 DFT data points to predict valence band maximum (VBM) levels relative to the vacuum level for arbitrary compositions $\{x, y, z\}$ . The model was validated with cross-validation mean absolute errors (MAE) of 4.8–12.3 meV.
Prediction Scope: The trained ML model was applied to a dense compositional grid (9,261 solid solutions) to map VBM and conduction band minimum (CBM) positions for the two most stable surface terminations, (010) and (011). CBM positions were derived by combining predicted VBM values with band gap predictions from a previous study.

Key Results

Tunability via Composition and Facet: The study reveals that band edges can be tuned by more than 1 eV through chemical composition alone. Crucially, for a fixed composition, the VBM and CBM can shift by up to 0.6 eV between the (010) and (011) surface terminations, despite these surfaces having nearly degenerate formation energies. This establishes facet selection as a design parameter comparable to chemical substitution.
Orbital Origins: The shifts in band alignment are attributed to a rigid displacement of the electrostatic reference level (surface dipoles) rather than the introduction of new electronic states. The VBM is dominated by Ch p-states (S/Br) with Bi s+p contributions, while the CBM is dominated by Bi p-states.
Photocatalytic Applications:
- (010) Surface: A wide range of compositions is suitable for water splitting. Selenium-rich and antimony-rich compositions favor hydrogen evolution (HER), while bismuth-rich and sulfur-rich compositions favor oxygen evolution (OER). The (010) surface also supports nitrogen reduction to ammonia and carbon dioxide reduction to methane for specific stoichiometries.
- (011) Surface: In stark contrast, (011)-oriented solid solutions generally fail to straddle the redox potentials required for water splitting (HER and OER), rendering them unsuitable for unassisted water splitting despite their structural similarity to the (010) phase.
- Other Reactions: The study identifies specific compositions capable of driving hydrogen peroxide ( $H_2O_2$ ) generation, with the (011) surface showing a distinct window for selenium-rich compositions.
Photovoltaic Applications: The study evaluated the compatibility of MChX solid solutions with common charge-selective contact materials (CdS, $V_2O_5$ $V_{2} O_{5}$ , $MoSe_2$ $M o S e_{2}$ ).
- All three materials align effectively as hole-transport layers (HTL) for various regions of the (010) compositional space, offering flexibility in contact selection.
- However, none of the tested materials function as suitable electron-transport layers (ETL) for either surface orientation. This indicates that current standard contacts align only as hole-selective contacts, leaving electron extraction as an unresolved design challenge for these devices.

Significance and Claims
The paper claims to establish a generalizable, ML-accelerated framework for probing surface properties in emergent semiconductors, bridging the gap between atomistic modeling and device-level design. The primary significance lies in demonstrating that:

Facet Engineering is Critical: Surface termination is a powerful, independent variable for tuning band alignments, with the (010) and (011) surfaces offering fundamentally different catalytic and electronic behaviors.
Targeted Material Discovery: The study provides specific, thermodynamically stable compositions for driving five distinct fuel-forming photocatalytic reactions (H2, NH3, CH4, H2O2, O2) and identifies the current limitations in electron-selective contact alignment for photovoltaics.
Methodological Advancement: By combining DFT with ML, the work overcomes the computational barriers of exploring vast compositional spaces, offering a scalable approach for the design of solar cells and photocatalytic devices beyond the pnictogen chalcohalide family.

More like this