PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential

arXiv:2607.28000 · cond-mat.mtrl-sci, physics.comp-ph · Submitted 2026-07-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential".

Mira: Here's the gist of the paper. Accurate atomistic modelling of iron (Fe) oxidation requires a reliable interatomic potential, which necessitates an extensive and representative first-principles dataset for training the interatomic potential <ref:2607.28000#pg2> The Perturbation Augmented Space group structure Sampling (PASS) method is proposed to generate extensive and representative small-cell structures with less than 10 atoms to develop a transferable machine learning interatomic potential (MLIP) for the Fe-O system <ref:2607.28000#pg5> The PASS Workflow The central bottleneck in MLIP development for complex reactive materials is representative dataset construction, which is solved by introducing PASS, a symmetry-informed, perturbation-augmented, descriptor-based one-shot sampling strategy This method avoids iterative active learning exploration and allows an ACE MLIP trained on small-cell structures to transfer to large structures and complex reaction processes The essential ideas behind PASS are first generating a comprehensive configurational space and then selecting the most representative structures while simultaneously capturing both equilibrium and non-equilibrium configurations How the training signal is built The dataset generation step involves two consecutive actions: dataset generation and descriptor-based selection Initial structures are obtained by stepwise relaxation of space group structures, followed by the application of rattle and stretch perturbations to introduce non-equilibrium LAEs These perturbed structures are then encoded using high-dimensional SOAP descriptors, projected into a 2D embedding space by principal component analysis (PCA), and selected by farthest point sampling (FPS) to retain the most representative structures The final training dataset is composed of initial and perturbed structures, including O data from existing datasets like OQMD39 and OMat2441 Where the gains come from The effectiveness of PASS is attributed to generating a broad range of LAEs, where the perturbation samples non-equilibrium configurations relevant to complex structures such as defects and interfaces The final dataset is designed to balance diversity and efficiency by retaining representative LAEs without relying on manually designed oxidation pathways or iterative active learning exploration cycles This systematic approach enables the ACE MLIP to capture both thermodynamics and kinetic properties across Fe and Fe-O systems Systematic Validation of the ACE MLIP for pure Fe The validation process systematically covers structural, mechanical, and chemical properties by starting from pure Fe and progressing to the Fe-O system The ACE MLIP yields mean absolute training errors of 31.89 meV/atom and 120.52 meV/Å for energy and force, respectively Validation tests include comparing the ACE MLIP predictions with DFT reference data for equation of state curves, lattice parameters, Bain path energy profiles, and generalized stacking fault energy profiles The ACE MLIP can well predict the energetics of various defects like vacancies and SIA formation energies, showing matching qualitative trends with DFT Systematic validation from thermodynamics to kinetics for Fe-O The ACE MLIP is used to predict the structural properties of common Fe oxides, namely FeO, Fe2O3, and Fe3O4, describing their lattice constants and formation enthalpies with accuracy comparable to DFT (+U) results Furthermore, the ACE MLIP is applied to study large-scale BCC Fe surface oxidation simulations using a 1 ns high-temperature oxidation of the BCC Fe (100) surface The simulation demonstrates that the ACE MLIP can simultaneously capture the formation of oxides and oxide/Fe interfaces, O diffusion into the Fe bulk, and high-temperature Fe dynamics The results show that the dominant local structure of the oxide layer is FeO-like under simulated high-temperature conditions Discussion PASS generates a training dataset comprising small-cell structures in a one-shot way, avoiding the heavy computational cost of extending datasets using iterative active learning explorations The effectiveness of PASS can be attributed to generating a broad range of LAEs and sampling non-equilibrium configurations relevant to complex structures and critical processes It is noted that the current ACE MLIP does not explicitly consider magnetism, so it should be interpreted as a structurally transferable and computationally practical interatomic potential trained on a highly systematic, oxidation-agnostic, small-cell dataset under a fixed magnetic setting Future work could incorporate magnetic descriptors with explicit spin degrees of freedom or separate magnetic treatment of oxidation states The dataset generation involves using the pyiron integrated development environment (IDE) to generate initial space group structures and apply rattle and stretch perturbations using the ASSYST tool The LAEs are encoded using the SOAP descriptor, with specific parameters used for constructing the descriptor DFT calculations are performed using VASP with GGA in the PBE form, employing a k-point mesh and a cutoff energy of 520 eV The ACE MLIP is trained by employing the PACEMAKER package, utilizing a Finnis-Sinclair-type non-linear embedding function and optimizing parameters using the BFGS algorithm MD simulations are performed using LAMMPS, simulating O diffusion in Fe bulk and large-scale BCC Fe surface oxidation at 873 K for 1 ns Data availability Data will be made available upon reasonable request The code for the ACE MLIP will be publicly available once the manuscript is accepted The authors acknowledge support from project number T23003 and the DelftBlue supercomputer provided by the Delft High Performance Computing Center The research was carried out under project number T23003 in the framework of the Research Program of the Materials innovation institute (M2i) supported by the Dutch government References The references cited include works on iron oxidation, machine learning interatomic potentials, and specific DFT/MLIP methods for validation The list includes foundational texts like Lawless's work on the oxidation of metals and various papers detailing ReaxFF and MLIP development Specific references validate structural properties, defect energetics, and kinetic barriers in pure Fe systems References also cover the application of MLIPs to complex systems like Fe-O oxides and interfaces The work builds upon prior strategies such as ASSYST and RSS while introducing the PASS method for systematic, symmetry-informed sampling The references also include details on the computational methods used, such as VASP for DFT and LAMMPS for MD simulations The work draws on established methodologies in materials science to ensure the transferability of the MLIP across different material classes The paper concludes by demonstrating that PASS is capable of enabling accurate and transferable MLIPs for chemically complex reactive processes The authors declare no competing interests Page 33 ---

Title and authors: Kai: So we're diving into this paper by Zixiong Wei and Poulumi Dey about this PASS method for training a machine learning interatomic potential for iron oxygen. It sounds like they're tackling a really tough problem because modeling iron oxidation needs a super reliable interatomic potential, which means you need a huge, representative dataset.

Mira: Exactly. The paper suggests that the existing methods struggle with the structural and magnetic complexity of the iron oxygen system, making it hard to build good training sets for these potentials. They are proposing this PASS method to generate extensive and representative small-cell structures with less than ten atoms.

Lev: So they're focusing on building a transferable machine learning interatomic potential specifically for the Fe oxygen system using these tiny structures as their foundation. That sounds like a smart way to avoid the massive computational expense of generating huge datasets through traditional methods like iterative active learning exploration, which is what they call the central bottleneck in MLIP development.

Kai: They introduce PASS as this symmetry-informed, perturbation-augmented, descriptor-based one shot sampling strategy to solve that data generation problem. The idea is to generate a broad range of small cell structures while simultaneously capturing both equilibrium and non equilibrium configurations, which is something they think is really important for understanding complex stuff like defects and interfaces.

Mira: And how they build that training signal involves two steps: first generating the initial structures through stepwise relaxation of space group structures, then applying rattle and stretch perturbations to introduce those non-equilibrium lattice energy states. Then, these perturbed structures are encoded using high dimensional SOAP descriptors, projected into a two dimensional embedding space by principal component analysis, and selected by farthest point sampling to keep the most representative ones.

Lev: That sounds like a systematic way to ensure you get diversity without spending endless time manually designing oxidation pathways or running those iterative exploration cycles they mentioned. But I'm wondering how well this one shot approach actually transfers when we move from these small cells to the much larger structures we care about in real simulations, like interfaces or large scale oxidation processes.

Title and authors: Kai: They show that this method allows an ACE machine learning interatomic potential trained on these small cell structures to transfer effectively to large structures and complex reaction processes, which is a big deal for making it practical. The paper aims to capture both the thermodynamics and kinetic properties across the iron and iron oxygen systems using this approach.

Mira: They validate this by starting with pure iron, then moving onto the Fe oxygen system, showing that the ACE machine learning interatomic potential yields mean absolute training errors of thirty-one point eight nine meV per atom for energy and one hundred twenty point five two meV per angstrom for force when compared to DFT reference data.

Lev: That error level sounds promising for a potential, but I'm looking at the validation tests they run, like comparing the ACE machine learning interatomic potential predictions with DFT data for equation of state curves, lattice parameters, and Bain path energy profiles. The paper claims it can well predict the energetics of various defects like vacancies and SIA formation energies while showing matching qualitative trends with DFT results.

Kai: And they don't stop there; they apply this ACE machine learning interatomic potential to predict the structural properties of common iron oxides like FeO, Fe2O3, and Fe3O4, describing their lattice constants and formation enthalpies with accuracy comparable to DFT plus U results.

Mira: Furthermore, for kinetic studies, they use this ACE machine learning interatomic potential to study large scale BCC iron surface oxidation simulations using a one nanosecond high temperature oxidation of the BCC iron ten zero surface. The simulation demonstrates that the ACE machine learning interatomic potential can capture the formation of oxides and oxide iron interfaces, oxygen diffusion into the iron bulk, and high temperature iron dynamics simultaneously.

Lev: So they’re showing that this method can handle both the thermodynamics and kinetics at once in a large scale simulation setup, which is where we really need it for understanding things like surface reconstruction or diffusion barriers. They also show that under simulated high temperature conditions, the dominant local structure of the oxide layer is FeO like.

Title and authors: Kai: The discussion points out that PASS generates a training dataset in a one shot way, avoiding the heavy computational cost of extending datasets using iterative active learning explorations. The effectiveness comes from generating a broad range of lattice energy states and sampling non equilibrium configurations relevant to complex structures and critical processes.

Mira: They also point out that the current ACE machine learning interatomic potential does not explicitly consider magnetism, so it should be interpreted as a structurally transferable and computationally practical interatomic potential trained on a highly systematic, oxidation agnostic small cell dataset under a fixed magnetic setting. Future work could incorporate magnetic descriptors with explicit spin degrees of freedom or separate magnetic treatment of oxidation states.

Lev: I think that’s the necessary caveat; if you're building real hardware, you need to know what assumptions you're making about the underlying physics, and ignoring magnetism means this is currently structural transferability more than a fully magnetic model. It’s a practical starting point for simulation work.

Kai: So to wrap up on this PASS method, it generates training data systematically and efficiently without needing those time consuming iterative cycles, allowing the ACE machine learning interatomic potential to capture both thermodynamics and kinetics across the iron and iron oxygen systems in a very systematic way.

Mira: It’s an interesting way to bridge the gap between complex real world chemistry and computationally feasible machine learning models for materials science. It gives us a tool that can handle both structural properties of pure iron and the more complex behavior of oxides.

Lev: For me, it means we have a much more robust way to generate reliable force fields, even if they are structurally transferable rather than fully magnetic yet. The systematic approach is what really matters for building tools that can actually run on real hardware efficiently.

Kai: That’s all for this paper by Zixiong Wei and Poulumi Dey on PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential. Next up we have a new look at avalanche criticality emerging from thermal fluctuation in a quiescent glass.

The paper's summary: Kai: So, this paper is about this PASS method for training an interatomic potential for iron oxygen that uses small cell structures to learn how to predict bigger things like oxidation or defects without needing a massive dataset from scratch.

Mira: Exactly, it’s about building a machine learning interatomic potential for iron oxygen by using a systematic sampling strategy called PASS, which lets them create representative small-cell structures in one shot instead of doing all that slow, iterative exploration.

Lev: What this means is they're trying to solve the data generation bottleneck for MLIP development so you don't have to spend forever on active learning cycles just to get a training set ready for complex materials.

Kai: The core idea is that by using symmetry and adding small perturbations, they can sample a really broad range of structures, including those non-equilibrium ones relevant to defects and interfaces.

Mira: And they do this by first relaxing the space group structures and then applying some rattle and stretch perturbations to introduce those non-equilibrium lattice energy states, which is crucial for capturing real material behavior.

Lev: From an engineering standpoint, that one shot sampling strategy is smart because it bypasses the heavy computational cost of building massive datasets through traditional methods.

Kai: They then encode these structures using SOAP descriptors, project them into a two dimensional embedding space with principal component analysis, and select the most representative ones using farthest point sampling.

Mira: That means they are systematically capturing both equilibrium and non-equilibrium configurations simultaneously, which helps the resulting potential be more transferable across different scales.

Lev: The results show this ACE machine learning interatomic potential can predict structural properties of common iron oxides like FeO, Fe2O3, and Fe3O4 with accuracy that’s comparable to doing density functional theory with a plus U correction.

Kai: They also tested its kinetic capabilities by using it to study large scale BCC iron surface oxidation simulations at high temperature and found it captures the formation of oxides and interfaces along with oxygen diffusion into the iron bulk.

Mira: The validation tests also showed that the potential can predict vacancy migration barriers in FeO and Fe3O4 with a trend matching density functional theory results, which is a big deal for kinetics.

Lev: But there’s still one thing they flagged—the current ACE machine learning interatomic potential doesn't explicitly consider magnetism, so you should think of it as being structurally transferable under a fixed magnetic setting.

Kai: That’s the caveat; it’s a highly systematic and oxidation agnostic approach for building a practical potential right now, but incorporating magnetism with explicit spin degrees of freedom is definitely what they need to tackle next.

Mira: So, this PASS method gives us a systematic way to generate training data efficiently, allowing the ACE machine learning interatomic potential to capture both thermodynamics and kinetics across the iron and iron oxygen systems in a very practical manner.

Lev: It’s a solid foundation for building tools that can handle complex reactive processes without needing those slow iterative exploration cycles every single time you want to test something new.

Kai: We're seeing a clear path here for making these MLIPs more accessible for simulating real world chemistry, moving beyond just the purest systems we have now.

Mira: It’s interesting because they are showing that systematic generation can deliver results comparable to much more expensive, traditional methods in terms of both structural and kinetic predictions.

Lev: Next up, we're going to look at how avalanche criticality emerges from thermal fluctuation in a quiescent glass.

The paper's improvements: Kai: So, this part is about what they think we can do better with this PASS method for training those iron oxygen potentials because they are looking at how to refine it further.

Mira: They suggest that the real improvement lies in making sure you're generating a really broad range of lattice energy states, especially the non-equilibrium ones, which is where the accuracy really comes from.

Lev: It sounds like they want to push for even more diversity in those perturbed structures to ensure the final potential has wider applicability across different chemical environments.

Kai: They are focusing on generating a training dataset that balances diversity and efficiency by retaining representative lattice energy states without relying on manually designed oxidation pathways or those slow iterative exploration cycles.

Mira: That's smart because it means the ACE machine learning interatomic potential will be more robust when you try to use it for things like complex defect modeling or interfaces.

Lev: If they can get a really good balance between diversity and efficiency in that one shot sampling, that translates directly into a faster path to having a reliable model ready for real hardware testing.

Kai: They also point out the limitations of the current setup—that it doesn't explicitly consider magnetism—so they suggest future work could incorporate magnetic descriptors with explicit spin degrees of freedom or separate magnetic treatment of oxidation states.

Mira: That makes sense; if you’re modeling things like FeO or Fe3O4, ignoring magnetism means you’re missing some key physics that dictates how those materials actually behave under different conditions.

Lev: For someone building hardware, that separation is important because spin dynamics can change things dramatically in a simulation setup and add significant overhead to the computational cost.

Kai: So they are essentially saying that while this method is great for structural transferability, the next step for this potential should be to integrate magnetic information more deeply into the descriptor space.

Mira: It’s about moving from a purely structural model to one that understands how magnetism influences those iron oxide properties, which is where a lot of the interesting physics lies.

Lev: If they can do that, it opens up new avenues for simulating things like surface reconstructions or complex magnetic ordering in these systems.

Kai: So, the next phase for this research seems to be integrating magnetic descriptors to build an interatomic potential that understands both the structure and the magnetism of iron oxygen materials.

Mira: It’s about capturing a richer physics picture, which is essential when you want to simulate things beyond just basic structural properties.

Lev: We're moving towards a more comprehensive model that can handle those magnetic effects without incurring too much computational penalty on the actual simulation runs.

Kai: It sounds like they're building a potential that’s not just structurally sound, but also physically complete for the iron system as a whole.

Conclusion: Kai: So we're wrapping up on "PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential." Essentially, this paper shows how to build a training dataset for an iron oxygen potential in one shot using a systematic sampling method that avoids those time-consuming iterative exploration cycles.

Mira: Right, it's about getting high quality training data efficiently by smartly sampling both equilibrium and non-equilibrium configurations through symmetry and perturbation techniques.

Lev: What this means for the hardware side is that we get a more reliable potential faster, which is crucial when we’re trying to simulate things like defect migration or diffusion barriers on real systems.

Kai: It really gives us a systematic way to bridge the gap between complex chemistry and computationally feasible machine learning models for materials science.

Mira: The implication is that we can now predict structural properties of common iron oxides like FeO and Fe2O3 with accuracy comparable to density functional theory plus U results.

Lev: And they've shown it can handle kinetic stuff too, like predicting vacancy migration barriers, which means the potential is useful for simulating reaction rates.

Kai: They also highlighted that the ACE machine learning interatomic potential doesn't explicitly consider magnetism right now, so you should interpret it as a structurally transferable and computationally practical potential under a fixed magnetic setting.

Mira: That’s the caveat; it’s not fully magnetic yet, but it’s a great starting point because it’s systematic and oxidation agnostic for building those small-cell training sets.

Lev: For us, that means we have a solid foundation that we can definitely expand later by incorporating spin degrees of freedom without having to rebuild the entire dataset from scratch.

Kai: So this whole paper is about using PASS to create a robust machine learning interatomic potential for iron oxygen by generating diverse training data systematically and efficiently.

Mira: It shows that this systematic approach can capture both thermodynamics and kinetics across the iron and iron oxygen systems in a very practical way.

Lev: It’s a big step toward making MLIPs more accessible for simulating complex reactive processes without needing those slow iterative exploration cycles every time you want to test something new.

Kai: We’re seeing a clear path here for making these MLIPs more accessible for simulating real world chemistry, moving beyond just the purest systems we have now.

Mira: It’s interesting because they are showing that systematic generation can deliver results comparable to much more expensive, traditional methods in terms of both structural and kinetic predictions.

Lev: We're moving towards a more comprehensive model that can handle those magnetic effects without incurring too much computational penalty on the actual simulation runs.

Kai: So this whole paper is about using PASS to create a robust machine learning interatomic potential for iron oxygen by generating diverse training data systematically and efficiently.

Mira: It shows that this systematic approach can capture both thermodynamics and kinetics across the iron and iron oxygen systems in a very practical way.

Lev: It’s a big step toward making MLIPs more accessible for simulating complex reactive processes without needing those slow iterative exploration cycles every time you want to test something new.

Kai: We’re seeing a clear path here for making these MLIPs more accessible for simulating real world chemistry, moving beyond just the purest systems we have now.

Mira: It’s interesting because they are showing that systematic generation can deliver results comparable to much more expensive, traditional methods in terms of both structural and kinetic predictions.

Lev: We're moving towards a more comprehensive model that can handle those magnetic effects without incurring too much computational penalty on the actual simulation runs.

Kai: That’s all for "PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential." It’s a solid piece of work on systematic data generation.

Mira: It really proves that targeted sampling can be just as effective as brute force when you're looking to build these potentials.

Lev: Next up, we're going to look at how avalanche criticality emerges from thermal fluctuation in a quiescent glass.

Zixiong Wei, Fei Shuang, Poulumi Dey

Department of Materials Science and Engineering, Faculty of Mechanical Engineering, Delft University of Technology · State Key Laboratory of Nonlinear Mechanics, Institute of Mechanics, Chinese Academy of Sciences

cond-mat.mtrl-sci, physics.comp-ph

Submitted: 2026-07-30

Updated: 2026-10-30

Comments: 39 pages, 8 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 80/100

The gist: Accurate atomistic modelling of iron (Fe) oxidation requires a reliable interatomic potential, which necessitates an extensive and representative first-principles dataset for training the interatomic…

Key concepts

PASS Method
A systematic sampling strategy that uses symmetry information and small perturbations to generate a wide variety of atomic structures. It creates a diverse training set in a single step, bypassing the need for slow, iterative learning cycles typical in other MLIP development methods.
SOAP Descriptors
High-dimensional mathematical descriptions used to encode the local atomic environment around each atom. These descriptors capture detailed structural information necessary for the machine learning model to learn complex interactions accurately during training.
ACE MLIP
The resulting machine learning interatomic potential trained using the PASS dataset. It is designed to be transferable, meaning it can accurately predict physical properties like energy and forces across different scales, from pure iron to complex iron oxides.

Terminology

Summary

Accurate atomistic modelling of iron (Fe) oxidation requires a reliable interatomic potential, which necessitates an extensive and representative first-principles dataset for training the interatomic potential. p. 0

Where the paper says it

Page 1 of 33 PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential Zixiong Wei1 , Fei Shuang1, 2 , Poulumi Dey1* 1 Department of Materials Science and Engineering, Faculty of Mechanical Engineering, Delft University of Technology, Mekelweg 2, 2628 CD Delft, The Netherlands 2 State Key Laboratory of Nonlinear Mechanics, Institute of Mechanics, Chinese Academy of Sciences, Beijing 100190, China Abstract: Accurate atomistic modelling of iron (Fe) oxidation requires a reliable interatomic potential, which necessitates an extensive and representative first-principles dataset for training the interatomic potential. However, Fe-oxygen (O) system is known for its structural and magnetic complexity, rendering the generation of high-quality dataset challenging.

This work proposes the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms. p. 0
Where the paper says it

However, Fe-oxygen (O) system is known for its structural and magnetic complexity, rendering the generation of high-quality dataset challenging. In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms.

They present a systematic approach to developing a first of its kind transferable machine learning interatomic potential (MLIP) for Fe-O system based on the atomic cluster expansion (ACE) framework. p. 0
Where the paper says it

In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms. We present a systematic approach to developing a first of its kind transferable machine learning interatomic potential (MLIP) for Fe-O system based on the atomic cluster expansion (ACE) framework.

The PASS method yields an accurate and transferable MLIP which is capable of capturing the reactive complexity of oxide growth while remaining computationally practical for extended systems. p. 0
Where the paper says it

We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP. This work demonstrates that the PASS method yields an accurate and transferable MLIP which is capable of capturing the reactive complexity of oxide growth while remaining computationally practical for extended systems.

Pruning redundant reasoning tokens

The central bottleneck in MLIP development for complex reactive materials is representative dataset construction. The PASS workflow involves two consecutive steps, namely dataset generation and descriptor-based selection. p. 4

Where the paper says it

The essential ideas behind PASS are first generating a comprehensive configurational space and then selecting the most representative structures while simultaneously capturing both equilibrium and non-equilibrium configurations. Specifically, two consecutive steps are performed, namely dataset generation and descriptor-based selection, as shown in Fig.

The initial space group structures are generated using PyXtal and then optimized through stepwise volume and full structural relaxation to obtain initial equilibrium configurations. p. 5
Where the paper says it

(a) Dataset generation. Space group structures are first generated using PyXtal and then optimized through stepwise volume and full structural relaxation to obtain initial equilibrium configurations.

Subsequently, the volumes of PyXtal structures are relaxed, leading to Volmin structures, followed by Fullmin structures. p. 6
Where the paper says it

Subsequently, the volumes of PyXtal structures are relaxed, leading to Volmin structures. Afterwards, Volmin structures are fully relaxed to form Fullmin structures.

To generate perturbed structures, two separate actions are taken on Fullmin structures, namely rattle and stretch. p. 6
Where the paper says it

, PyXtal, Volmin, and Fullmin) are collected, forming the initial set of structures for the final high-fidelity DFT labelling. Next, to generate perturbed structures, two separate actions are taken on Fullmin structures, namely rattle and stretch.

The first crucial element in the PASS method is to extensively sample the configurational space in a controlled way by applying a grid of perturbation parameters. p. 6
Where the paper says it

The first crucial element in the PASS method is to extensively sample the configurational space in a controlled way. Specifically, we apply a grid of perturbation parameters (i.

How the training signal is built

The generated perturbed structures are encoded using high-dimensional SOAP descriptors, projected into a 2D embedding space by principal component analysis (PCA), and selected by farthest point sampling (FPS) to retain the most representative structures while reducing redundancy. The composition of the final training dataset is shown in Table 1, which includes initial and perturbed structures for Fe and Fe-O systems. p. 7

Where the paper says it

The composition of the final training dataset is shown in Table 1. Note that the PASS method is only applied to obtain Fe and Fe-O structures following the procedures described above.

The purpose of PASS is to construct a compact and representative one that spans a broad range of physically relevant LAEs. p. 7
Where the paper says it

Composition of the final training dataset. System Subset Number of structures Fe Initial 3,063 Perturbed 10,000 Fe-O Initial 3,802 Perturbed 10,000 O OQMD 33 OMat24 1,193 Total 28,091 Generation of extensive and representative dataset The purpose of PASS is not simply to enlarge the training dataset, but to construct a compact and representative one that spans a broad range of physically relevant LAEs.

The initial space group structures provide chemically and crystallographically diverse equilibrium configurations, while the systematic and extensive rattle and stretch perturbations introduce non-equilibrium LAEs potentially associated with defective, strained, and transition state configurations. p. 7
Where the paper says it

System Subset Number of structures Fe Initial 3,063 Perturbed 10,000 Fe-O Initial 3,802 Perturbed 10,000 O OQMD 33 OMat24 1,193 Total 28,091 Generation of extensive and representative dataset The purpose of PASS is not simply to enlarge the training dataset, but to construct a compact and representative one that spans a broad range of physically relevant LAEs. The initial space group structures provide chemically and crystallographically diverse equilibrium configurations, while the systematic and extensive rattle and stretch perturbations introduce non-equilibrium LAEs potentially associated with defective, strained, and transition state configurations.

Where the gains come from

The final dataset is designed to balance diversity and efficiency such that it retains representative LAEs without relying on manually designed oxidation pathways or iterative active learning exploration cycles. The small-cell PASS dataset covers the LAE space sampled by these application-specific datasets, demonstrating its outstanding representativeness. p. 8

Where the paper says it

The small-cell PASS dataset covers the LAE space sampled by these application-specific (e. , deformation and fracture) datasets.

For the Fe-O system, the LAEs in SPG Perturbed dataset not just cover the configurational space of all other datasets but also extend beyond it, owing to the extensive perturbations. p. 7
Where the paper says it

System Subset Number of structures Fe Initial 3,063 Perturbed 10,000 Fe-O Initial 3,802 Perturbed 10,000 O OQMD 33 OMat24 1,193 Total 28,091 Generation of extensive and representative dataset The purpose of PASS is not simply to enlarge the training dataset, but to construct a compact and representative one that spans a broad range of physically relevant LAEs. The initial space group structures provide chemically and crystallographically diverse equilibrium configurations, while the systematic and extensive rattle and stretch perturbations introduce non-equilibrium LAEs potentially associated with defective, strained, and transition state configurations.

Systematic validation of the ACE MLIP for pure Fe

The ACE MLIP yields the mean absolute training errors of 31.89 meV/atom and 120. p. 2

Where the paper says it

Instead, they use atomic descriptors or structural fingerprints to represent or encode the local atomic environments (LAEs). A number of atomic descriptors have emerged to capture the essential characteristics of LAEs, including atom-centered symmetry functions26, smooth overlap of atomic positions (SOAP)27, moment tensor28, and atomic cluster expansion (ACE)29.

52 meV/Å for energy and force, respectively. p. 6
Where the paper says it

Afterwards, Volmin structures are fully relaxed to form Fullmin structures. Note that for the purpose of obtaining equilibrium configurations in a less computational demanding way, the stepwise relaxation processes are performed at a lower density-functional theory (DFT) accuracy in terms of both energy and force convergence criteria and k-point density.

Validation tests are performed in a progressive manner, starting from thermodynamics to kinetics and from pure Fe to Fe-O system. p. 1
Where the paper says it

Notably, this challenge can be better resolved by atomistic modelling, especially molecular dynamics (MD) simulation, where structural and chemical information at the atomistic level, as well as thermodynamics and kinetics information can be obtained simultaneously. Nevertheless, accurately modelling Fe oxidation is a challenging task due to the complexity of Feoxygen (O) system.

The ACE MLIP can well capture the influence of temperature and accurately describe the lattice thermal expansion of both phases. p. 2
Where the paper says it

Instead, they use atomic descriptors or structural fingerprints to represent or encode the local atomic environments (LAEs). A number of atomic descriptors have emerged to capture the essential characteristics of LAEs, including atom-centered symmetry functions26, smooth overlap of atomic positions (SOAP)27, moment tensor28, and atomic cluster expansion (ACE)29.

Furthermore, it can distinguish shear deformations on the two planes, obtaining unstable (i. p. 6
Where the paper says it

Note that two strain modes are randomly selected i. , 70% hydrostatic (diagonal) and 30% shear (off-diagonal), with strain magnitudes 𝜖()*!

e., maximum) stacking fault energies close to the DFT results.

Systematic validation from thermodynamics to kinetics for Fe-O

The ACE MLIP can describe those typical Fe oxides’ structural properties as accurate as DFT (+U) results, with performance as good as ACE-Bienvenu. The ACE MLIP can capture the correct thermodynamics by exhibiting formation enthalpies that exhibit the same qualitative trend as the DFT (+U) results. p. 0

Where the paper says it

We thoroughly validate the accuracy and capability of the ACE MLIP across both pure Fe and Fe-O systems through bulk, surface, and interface properties. We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP.

The ACE MLIP performs similarly as the ACE-Bienvenu and shows a comparable trend with respect to DFT for vacancy diffusion barriers in FeO and Fe3O4. p. 4
Where the paper says it

We solve this by introducing PASS, a symmetry-informed, perturbation-augmented, descriptor-based one-shot sampling strategy. PASS not only avoids iterative active learning exploration, but also allows an ACE MLIP trained on small-cell structures to transfer to: (1) large structures, such as surfaces and interfaces; (2) application-relevant events, for example atomic diffusion and defect migration; and (3) complex reaction processes, for instance large-scale Fe oxidation.

Application to large-scale oxidation simulation

The ACE MLIP is further applied to study large-scale BCC Fe surface oxidation. The simulation demonstrates that the ACE MLIP can simultaneously capture the formation of oxides and oxide/Fe interfaces, O diffusion into the Fe bulk, and high-temperature Fe dynamics. p. 0

Where the paper says it

We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP. This work demonstrates that the PASS method yields an accurate and transferable MLIP which is capable of capturing the reactive complexity of oxide growth while remaining computationally practical for extended systems.

The layer-resolved composition profile shows that the oxide layer region approaches an Fe to O atomic ratio close to that of FeO. p. 2
Where the paper says it

Consequently, a reliable, robust and transferable interatomic potential is imperative, i. , the one which is capable not only of capturing the elementary reaction steps during initial oxidation but also of distinguishing between the nucleation, growth and phase transformations among the various iron oxide phases during growth of the oxide layer.

The RDF of the extracted oxide layer reproduces the main Fe-O, Fe-Fe, and O-O features of DFT-relaxed FeO structure. p. 0
Where the paper says it

We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP. This work demonstrates that the PASS method yields an accurate and transferable MLIP which is capable of capturing the reactive complexity of oxide growth while remaining computationally practical for extended systems.

Discussion

PASS generates training dataset comprising of small-cell structures in the one-shot way. This avoids the heavy computational cost of extending dataset using iterative active learning explorations. p. 4

Where the paper says it

We solve this by introducing PASS, a symmetry-informed, perturbation-augmented, descriptor-based one-shot sampling strategy. PASS not only avoids iterative active learning exploration, but also allows an ACE MLIP trained on small-cell structures to transfer to: (1) large structures, such as surfaces and interfaces; (2) application-relevant events, for example atomic diffusion and defect migration; and (3) complex reaction processes, for instance large-scale Fe oxidation.

The effectiveness of PASS can be attributed to the fact that it generates a broad range of LAEs, and the perturbation samples non-equilibrium configurations relevant to complex structures (e. p. 7
Where the paper says it

System Subset Number of structures Fe Initial 3,063 Perturbed 10,000 Fe-O Initial 3,802 Perturbed 10,000 O OQMD 33 OMat24 1,193 Total 28,091 Generation of extensive and representative dataset The purpose of PASS is not simply to enlarge the training dataset, but to construct a compact and representative one that spans a broad range of physically relevant LAEs. The initial space group structures provide chemically and crystallographically diverse equilibrium configurations, while the systematic and extensive rattle and stretch perturbations introduce non-equilibrium LAEs potentially associated with defective, strained, and transition state configurations.

g., defects and interfaces) and critical processes (e. p. 8
Where the paper says it

Each point represents the LAE of an individual atom. The broad coverage indicates that the smallcell PASS dataset contains diverse LAEs relevant to bulk oxides, surfaces, defects, and interfaces.

g., adsorption and diffusion). p. 1
Where the paper says it

Therefore, Fe oxidation has been extensively investigated by various experimental techniques, such as low-energy electron diffraction6,7, X-ray scattering8 , and in situ environmental scanning/transmission electron microscopy9 . From an atomistic perspective, oxidation involves a series of critical microscopic events10, including atomic adsorption, diffusion, and surface reconstruction.

It is worthwhile noting that the current ACE MLIP does not explicitly consider magnetism and thus is not a spin-aware interatomic potential. Consequently, its applicability is limited in regimes where magnetism strongly influences energetics and kinetics. p. 1

Where the paper says it

The combination of those crucial events poses a challenge to experimental studies, that is how to achieve three-dimensional atomic resolution for structural and chemical information during real-time oxidation monitoring11. Notably, this challenge can be better resolved by atomistic modelling, especially molecular dynamics (MD) simulation, where structural and chemical information at the atomistic level, as well as thermodynamics and kinetics information can be obtained simultaneously.

Future work could incorporate magnetic descriptors with explicit spin degrees of freedom or separate magnetic treatment of oxidation states. p. 0
Where the paper says it

However, Fe-oxygen (O) system is known for its structural and magnetic complexity, rendering the generation of high-quality dataset challenging. In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms.

Methods

The pyiron integrated development environment (IDE) is adopted as the dataset generation platform. The initial space group structures are generated by the PyXtal package and then the rattle and stretch perturbations are applied to Fullmin structures using the ASSYST tool implemented in pyiron IDE. p. 5

Where the paper says it

Space group structures are first generated using PyXtal and then optimized through stepwise volume and full structural relaxation to obtain initial equilibrium configurations. Perturbed structures are subsequently generated from the relaxed structures through systematic rattle and stretch operations, introducing non-equilibrium LAEs.

To encode the LAEs, the SOAP descriptor is adopted as implemented in the DScribe package. p. 6
Where the paper says it

The second crucial element in PASS method lies in the descriptor-based selection or sampling process, as shown in Fig. After filtering out structures with unphysically short interatomic distances, high-dimensional SOAP descriptors are utilized to encode the perturbed structures’ LAEs.

The DFT calculations are performed using VASP at T = 0 K with GGA in the Perdew-Burke-Ernzerhof (PBE) form. p. 6
Where the paper says it

Afterwards, Volmin structures are fully relaxed to form Fullmin structures. Note that for the purpose of obtaining equilibrium configurations in a less computational demanding way, the stepwise relaxation processes are performed at a lower density-functional theory (DFT) accuracy in terms of both energy and force convergence criteria and k-point density.

MLIP training and validation

The ACE MLIP is trained by employing the PACEMAKER package, using a Finnis-Sinclair-type non-linear embedding function. The parameters are optimized by the Broyden-Fletcher-GoldfarbShanno (BFGS) algorithm for 4,000 steps. Most validation tests are conducted in a workflow based on Atomic Simulation Environment (ASE) package. p. 3

Where the paper says it

proposed the Automated Small SYmmetric Structure Training (ASSYST) strategy for constructing systematic training dataset, based on extensively exploring all possible space groups of crystal structures37,38. Its effectiveness was demonstrated by various tests on outof-distribution structures (i.

The MD simulations are performed using Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS). p. 0
Where the paper says it

We thoroughly validate the accuracy and capability of the ACE MLIP across both pure Fe and Fe-O systems through bulk, surface, and interface properties. We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP.

MD simulations

For the Fe surface oxidation simulation, BCC(100) surface is selected and a 12×12×24 supercell is created. The temperature of the entire system is regulated at a constant 873 K by Nosé-Hoover heat bath method for 1 ns. p. 0

Where the paper says it

However, Fe-oxygen (O) system is known for its structural and magnetic complexity, rendering the generation of high-quality dataset challenging. In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms.

The results suggest that the short-range ordering in the oxide layer is predominantly FeO-like under the simulated high-temperature conditions.

Data availability

Data will be made available upon reasonable request. The files for the ACE MLIP will be publicly available once the manuscript is accepted. p. 0

Where the paper says it

In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms. We present a systematic approach to developing a first of its kind transferable machine learning interatomic potential (MLIP) for Fe-O system based on the atomic cluster expansion (ACE) framework.

Code availability

The files for the ACE MLIP will be publicly available once the manuscript is accepted.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Systematic dataset generation for complex reactive systems: The PASS method provides a symmetry-informed, perturbation-augmented, descriptor-based one-shot sampling strategy for building compact and diverse MLIP training dataset. This avoids iterative active learning exploration cycles and allows the MLIP to learn from small cells to transfer to large systems.

  2. Transferability across system scales: The ACE MLIP trained on small cell structures can be applied to large structures, such as surfaces and interfaces; (2) application-relevant events, for example atomic diffusion and defect migration; and (3) complex reaction processes, for instance large-scale Fe oxidation. This enables simulation of large-scale BCC Fe surface oxidation while capturing the early-stage coupling between surface adsorption, diffusion, and oxide growth.

  3. Accurate prediction of oxide formation: The ACE MLIP can predict the structural properties of common Fe oxides like FeO, Fe2O3, and Fe3O4 as accurate as DFT (+U) results, demonstrating its ability to capture the formation enthalpies with a qualitative trend matching DFT.

  4. Kinetic barrier simulation: The ACE MLIP is validated for kinetic properties by calculating Fe vacancy migration barriers in FeO and Fe3O4 using the CINEB method, showing it can capture these barriers similarly to ACE-Bienvenu and with a comparable trend with respect to DFT.

  5. Structural disorder characterization: The MLIP simulation of large-scale oxidation results in an RDF for the oxide layer that reproduces the main Fe-O, Fe-Fe, and O-O peak positions of DFT-relaxed FeO structure, indicating the system captures structural disorder expected at finite temperatures.

Abstract

Iron (Fe) oxidation is a fundamental phenomenon that is scientifically vital and technologically significant. Despite substantial experimental efforts, accurate atomistic modelling of Fe oxidation remains elusive, which requires a reliable interatomic potential, whose development hinges on a representative first-principles training dataset. However, Fe-oxygen (O) system is known for its structural complexity, rendering the generation of high-quality first-principles dataset challenging. In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive dataset for Fe-O system consisting of small-cell structures (no more than 10 atoms). Based on atomic cluster expansion (ACE) framework, we present a systematic approach to developing a transferable machine learning interatomic potential (MLIP). The accuracy and capability of the ACE potential are thoroughly validated across pure Fe and Fe-O systems through bulk, surface, and interface properties. The large-scale Fe oxidation simulations with the ACE potential capture key features of the oxidation process, including the formation of an FeO-like structure consistent with recent experimental observations. This work demonstrates that the PASS method yields an accurate and transferable MLIP that can capture the reactive complexity of oxide growth while remaining computationally practical for extended systems.

Sources

Related papers