RAFFLE: active learning accelerated interface structure prediction

Posted on by

When building any device — from solar cells to batteries — you inevitably bring different materials together. Each of those joins is an interface, and understanding its atomic structure is critical for predicting how the device will perform.

The problem? Predicting interface structures at the atomic scale is notoriously difficult.

To tackle this, we have developed RAFFLE: an open-source library for automated interface structure prediction, designed to make the process faster, smarter, and more accessible.

📄 Read the full article (npj Computational Materials):
https://doi.org/10.1038/s41524-025-01749-5

💻 Get the code (also via pip):
https://github.com/ExeQuantCode/RAFFLE/

Artistic depiction of an atomic-scale interface between two crystals, illustrated with patterned bathroom tiles. The top purple-and-white diamond tiles are horizontally aligned, while the bottom lavender tiles with teal square and diamond patterns are diagonally aligned, symbolising crystal orientation and lattice mismatch. The top tile patterns morph into a ball-and-stick atomic structure, which then transitions into the patterns of the bottom tiles.

What RAFFLE Does

  • Automates interface structure prediction with minimal setup
  • Written in Python (with Fortran backend available)
  • Seamlessly integrates with ASE, so it slots easily into existing workflows
  • Generates diverse candidate interfaces from a given host structure
  • Uses active learning: iteratively feeding back energies and structures to refine predictions of the potential energy landscape
  • Compatible with any self-consistent calculator — from DFT to empirical potentials to ML interatomic models (e.g. MACE-MP-0, CHGNet)
  • Supports both stochastic and deterministic structure generation

Why It Matters

This work builds on our earlier demonstration that RAFFLE can predict the phase stability of interfaces (Phys. Rev. Lett. 132, 066201). With this release, RAFFLE is now a practical tool for:

  • Structure search of interface reconstructions
  • High-throughput studies of material interfaces
  • Exploring entirely new classes of interface-stabilised materials

Compatibility

RAFFLE is a Fortran library provided with a Python wrapper. As such, RAFFLE can be called as either a (most supported) Python library, a Fortran library, or a Fortran executable with an input file.

Additionally, ARTEMIS (the group’s abrupt interface generation tool) is being redeveloped to include a Python wrapper to enable a seamless workflow between the two packages. ARTEMIS can be used to generate host structures that can be fed into RAFFLE for subsequent structure search.

ARTEMIS currently supports Python library use, but it still under development, so bugs and crashes are known to occur. Please report them as issues in the GitHub issue tracker so that they can be addressed by the group.

Instability of 𝐴⁢𝐵⁢O3 perovskite surfaces induced by vacancy formation (𝐴=Ca,Sr,Ba; 𝐵=Ti,Zr,Sn)

Posted on by

We have recently published our work exploring the effects of vacancies on the stability of surfaces for a set of oxide perovskites.

Article: https://doi.org/10.1103/w683-tvc4

Within the article, we use density functional theory (DFT) to explore energetics of surfaces, vacancies, and charged defects. Energetics are calculated using the PBE GGA functional, whilst charge states are corrected for using the HSE06 hybrid functional. Vacancy and surface energetics are modelled for nine ABO3 perovskites (A=Ca,Sr,Ba; B=Ti,Zr,Sn), whilst charged defects are studied for SrSnO3 only.

We explore the relative energetic stability of two (001) oxide perovskite (ABO3) surfaces and identify that, theoretically, the AO surface is more energetically favourable than the BO2 surface. However, through exploration of A-site, B-site, and O-site vacancy formation as a function of depth, we identify that the AO surface is more prone to vacancy formation, thus reducing its long-term stability.

Modelling of charged vacancies within SrSnO3 bulk and slabs (as a function of distance from the surface) highlights the further instability of the AO surface due to vacancies, especially under oxygen rich environments.

RAFFLE – predicting interface structure

Posted on by

We have recently been working on a method for predicting the atomic structure of interfaces. This approach uses random structure search methods informed by genetic algorithms to reduce the intense computational cost typically associated with such methods. This method is called RAFFLE (pseudoRandom Approach For Finding Local Energetic minima)

Article: https://doi.org/10.1103/PhysRevLett.132.066201

Within the article, we showcase the capabilities of the RAFFLE method by applying it to predicting the phase of thin layers of magnesium oxide (MgO) when encapsulated between two layers of graphene. The results show that the rocksalt phase of MgO is heavily stabilised by encapsulation as compared to its other potential phases.

The RAFFLE methodology generates interface structures by taking a host structure and inserting new atoms based on three placement methods (where the specific method used for each atom is randomly selected, with weighting to specific methods pre-defined); these methods are 1) global minima identifier, 2) pseudorandom walk, and 3) void identifier. The particulars of methods 1 and 2 depend on utilising distribution functions that relate particular features in n-body distributions of existing known structures (with the same chemical composition) to energetic favourability (i.e. the distribution functions of each known structure are combined with weightings applied to each based on the structure’s formation energy).

Global minima identifier
This method involves discretising the unit cell into a grid and evaluating the energetic favourability of each point based on the existing distribution functions. The point with the lowest energy (i.e. most favourable) is selected for the next atom placement. This captures global energetic information of the system.

Pseudorandom walk
This method involves selecting a random point in the cell and then evaluating points within a certain radius of the point for more energetically favourable sites. If one is found, the same evaluation is then performed on that point. This process is repeated until no points within a defined radius are found to be more favourable. That final point is selected for the next atom placement. This captures atoms getting caught in local energetic minima within the system.

Void identifier
This method involves identifying the point in the unit cell with the lowest atomic density. This point is then selected for the next atom placement. This method captures grain seed sites and increasing entropy of a system, i.e. homogeneity.

The Hepplestone Research Group is currently working on a follow-up paper and the associated RAFFLE code. The authors hope to have the code open-source and publicly available within the coming months.

Batteries, batteries, batteries

Posted on by

Recently we have had a flurry of activity on battery materials, battery metamaterials and battery properties. I thought this was a good opportunity to briefly summarise our recent work.

We have worked with Andrew Hector and coworkers and Deregallera on Sn3N4 battery anodes for Na batteries. This experimentally realised material (made by Andrew and co) shows remarkable capacity and after the initial cycle shows a large capacity >200 mAhr/g over repeat cycling. It represents a much more realisable anode material for Na than other materials considered.



We have also done substantial work on examining the intercalation materials, the Transition Metal Dichalogenides (TMDC). Conor Price led an excellent work on characterising the entire set theoretically, using a clever stability metric to assess their intercalation capacity. A highlight of this work was ScS2, which shows considerable potential as a cathode, potentially better than the industry standard, NMC. This system still has significant room to grow as the alloying concepts that followed from LiCoO2 could easily be applied here.

Another output of this same work is a series of anodes which would be highly suitable for Mg batteries, and the ranges which they could be cycled in. Given the double valency of these systems, it shows a significant potential compared to Li ion.

A significant effort has been to explore the role of metamaterial concepts in battery materials. Two areas are of particular interest, the first is the classical layering, creating nanolaminates or superlattices and the second is macroscale patterning which has seen considerable early success in the battery community.

In the former of these, we have explored TMDC/graphene and TMDC/TMDC heterostructures (which has the voltages for MS2 shown above). As you can see, the voltages tend to sit in the average between the two constituent materials, and this we find that many of these properties can be well approximated through consideration of the equivalent property for the component layers. An additional example, is if the superlattice volumetric expansion were to be estimated by calculating the mean of the volumetric expansion arising in the component TMDCs, we could expect the result to deviate by up to a 2% error from what is observed in the actual superlattice, and the voltage profiles of the component materials provide bounds to the voltage profile exhibited by the constructed superlattice. Further, the unoccupied states of the host material are progressively filled with the addition of an intercalant, which follows the behaviour observed with the individual TMDCs. Most interestingly, the construction of superlattices allows for many improvements to component materials: formation of a superlattice can result in a reduction of the electronic band gap, hence improving electronic conductivity; conversion-resistant materials can be used to increase the stability of conversion-susceptible materials, extending their cyclability and lifetime; and materials can be chosen such that the overall voltage can be tuned towards specific values.

Congratulations to Joe Pitfield and Conor Price

Posted on by

Joe Pitfield (above) and Conor Price both passed their PhD viva’s this November after 3+ years (extensions due to COVID).

Joe’s PhD is on the exciting topic of interface structure prediction and our RAFFLE structure prediction method, which I will discuss in a future blog post. Its a been a real pleasure working with Joe on this and I am pleased with the outcome of his PhD. Joe has since taken a position at the Postdoctoral Researcher at the Department of Physics and Astronomy, University of Aarhus and continues to work on interface structure prediction with the Hammer group. Joe has also worked on battery materials and photocatalysts.

Conor has worked on many different projects during his PhD and has shown real adaptability. His main focus has been on how Transition Metal Dichalogenides can be used as battery materials. Using DFT, he has explored both single layer and superlattice structures and shown how the voltage and capacity for these systems changes depending on the stoichiometric make up. His leading article on this is available here: http://dx.doi.org/10.1039/d3ta00940h

Finally, I would be remiss to not mention the two of them worked together with Edward Baker on showing the potential of ScS2 as a cathode material, with gravimetric capacities comparable with NMC or the ilk. Personally, I think this is an outstanding result and I hope the community looks at this and examines this exciting cathode material. This article is available here: http://dx.doi.org/10.1039/d2cp05055b

Calcium-stannous oxide solid solutions for solar devices

Posted on by

Ned Taylor, Arnaldo Galbiati, Monica Saavedra, and Steve Hepplestone have just published an article exploring the potential of calcium-doped stannous oxide, (Sn:Ca)xO, for its potential as an active layer in a solar device. By identifying a suitable oxide active layer, the authors hope to design an all-oxide solar cell. This work was performed by Ned and Steve at the University of Exeter whilst working with Solaris Photonics.

In this article, the authors explore how doping stannous oxide, SnO, with calcium affects the electronic and optical properties. It is determined that a doping concentration of x=7:1 results in the most favourable properties for photovoltaic applications – a direct band gap of 1.5 eV. The study is expanded upon by exploring potential transport layers for this particular solid solution. CaO and TiO2 are identified as potential candidates for the hole and electron transport layers, respectively.

A potential all-oxide solar cell design is put forward by the authors, CaO/(Sn:Ca)7:1O/TiO2. It is hoped that this study grow new interest in all-oxide solar cells, which have been touted as potential replacements for current silicon solar cells due to their possible improved stability and efficiency, and reduced environmental and economic costs.

To find out more, follow the link to the article: https://doi.org/10.1063/5.0024947

ARTEMIS: Ab initio restructuring tool enabling the modelling of interface structures

Posted on by

Ned Taylor, Frank Davies, Isiah Rudkin, Conor Price, Ed Chan, and Steve Hepplestone have published an article detailing the group’s first large-scale commercial scientific software package, ARTEMIS. This work was led by Ned, Frank and Steve, with help from the entire ARTEMIS research group. Isiah aided in development of crucial modules and subroutines of the program during his Summer project with the group.

In this article, the authors detail the workflow of ARTEMIS, in addition to the methods and capabilities of its major subroutines: Lattice Matching, Surface Terminations, Interface Identification, Interface Shifting, and Intermixing. ARTEMIS can be used to generate a set of potential interfaces between any two given crystals, which are provided by the user. These structures can then be modelled using first principles or empirical modelling tools to identify the most energetically interface.

This software package has great potential to aid scientists in studying interface structures by reducing the time taken to explore then, as well as potentially removing human bias from the study. ARTEMIS identifies lattice matches within user-specified tolerances and shifts the two materials to compensate for missing bonds at the interface. To introduce the concept of diffusion, intermixing can be performed across the interface, which can relieve interface strain.

The software is freely available via this link: http://www.artemis-materials.co.uk

To find out more, follow the link to the article: https://doi.org/10.1016/j.cpc.2020.107515

The Potential of Overlayers on Tin-based Perovskites for Water Splitting

Posted on by

Ned Taylor, Conor Price, Alex Petkov, Marcus Carr, Jason Hale, and Steve Hepplestone have just published an article. The initial work was performed by Conor Price, Alex Petkov, Marcus Carr and Jason Hale during their Masters Physics course at the University of Exeter, with it then being expanded upon by Conor, Steve and Ned.

In this article, the authors explore the capabilities of tin-based oxide perovskites as photocatalysis. An investigation of their electron properties, as well as the reaction pathways associated with their surfaces, is presented. It is determined that SrSnO3 offers some potential as a photocatalyst.

The addition of an overlayer to the surface of the oxide perovskite SrSnO3 is then considered. It is determined that the inclusion of this thin surface coating leads to a drastic improvement of both the oxygen and the hydrogen evolution reactions. SrSnO3 with a ZrO2 overlayer is found to be capable of sustaining bifunctional water splitting at its surface (simultaneous hydrogen and oxygen gas production from water).

To find out more, here is a link to the article: https://doi.org/10.1021/acs.jpclett.0c00964

ARTEMIS Release

Posted on by

Good news Everyone!

Our project lead and lead developer have given the clear to to release the ARTEMIS code version 1.0.0 (Download from http://artemis-materials.co.uk/). For ARTEMIS enquires you can contact us at support@artemis-materials.co.uk. Also we have got our web page and wiki up and running.

Our lead developer, Ned, will be posting more about ARTEMIS soon!

leave a comment below.

The Fundamental Mechanism Behind Colossal Permittivity in Oxides

Posted on by

Ned Taylor, Frank Davies, Shane Davies, Conor Price, and Steve Hepplestone have published an article describing the atomic-scale mechanism that gives rise to colossal permittivity within samples of CaCu3Ti4O12 (CCTO). This work was conducted at the University of Exeter during Ned, Frank, Shane and Conor’s PhDs.

For two decades, experimental samples of CCTO have shown extremely high values of relative permittivity (typically on the order of 104). Thus far, it has been shown that such high permittivity values are not present in the bulk material, and that, instead, this phenomenon is caused by the formation of a strongly insulating material at the boundary between CCTO grains (which are characterised as semiconducting) – commonly termed as the internal barrier layer capacitance (IBLC).

In this article, the authors explore the origin of this phenomenon at the atomic scale in order to determine the exact cause of the IBLC. The authors identify the formation of a thin metallic region at the interface between the insulating grain boundaries and the semiconducting grains. This metallic layer could allow for a rapid dielectric response from the large grains, but prevent transport between grains, due to the insulating boundary; this manifests itself as a large dielectric response, or high permittivity, of the sample.

In understanding the mechanism behind this colossal permittivity, the capabilities and limits of this phenomenon can be better understood. This article can aid in the engineering of artificial systems with colossal permittivity.

To find out more, follow the link to the article: https://doi.org/10.1002/adma.201904746