AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D

arXiv:2603.27886 · cond-mat.mtrl-sci, cs.AI, physics.comp-ph · Submitted 2026-03-29 · 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: "AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D".

Mira: Here's the gist of the paper. Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate <ref:2603.27886#pg2> The paper presents Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity <ref:2603.27886#pg2>.

Kai: Framework Architecture Mat3ra-2D is organized as an implementation stack for practical generation of realistic structures, such as slabs, surfaces, interfaces, and defective configurations, with explicit provenance <ref:2603.27886#pg2> Rather than a single monolithic codebase, the methodology is distributed across standards, common data, implementation layers, and usage examples <ref:2603.27886#pg2>.

Mira: The components include: 1.

Kai: Standards package mat3ra-esse

24, 25: , which provides the schemas, ontology, categorization, and exchange format used throughout the ecosystem <ref:2603.27886#pg2> 2.

Paper summary: Kai: So to wrap up this section, we’ve seen that Mat3ra-2D is this open source framework for rapidly designing realistic two dimensional materials and related structures <ref:2603.27886#pg1,realistic two dimensional materials and related structures>. The core claim is that by focusing on how you build the structure—orientation, termination, defects—rather than just picking a material from a database, you can create datasets better suited for real world applications.

Mira: That’s the thesis. They argue that generating realistic structures isn't just selecting something from a database; it’s constructing and documenting the sequence of transformations that leads to the final geometry. This is especially acute for two dimensional and surface dominated systems where those creation choices really matter, like vacuum spacing or strain state.

Lev: The paper sets up a modular architecture with standards, data packages, an abstraction layer for code, and implementation packages all working together to handle these complex transformations in a way that preserves provenance.

Kai: They claim this combination of standards and transformation workflows is what allows them to address the need for realistic structures that reflect device modeling needs better.

Mira: And they show how they implement typical tasks, like constructing orientation specific slabs or strain matching interfaces, inside reusable Jupyter notebooks that function as both interactive documentation and templates for reproducible runs.

Lev: It’s about making this workflow explicit so you can see exactly what inputs led to what output, which is a big deal for debugging things later on when you move to actual hardware.

Kai: It sounds like the main point is moving from just having material data to having a traceable design process for creating the structures that matter.

Conclusion: Mira: So looking at the overall concept, "AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D" points to a system that acknowledges the limitations of current AI training data being too focused on ideal bulk crystals <ref:2603.27886#pg1,AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D>.

Kai: It means they’re trying to build a better way for AI models to learn from structures that actually mimic the complexity found in real devices, including surfaces and defects.

Lev: What this means for us in error correction is that if we can generate synthetic datasets with this level of structural detail, those models might be less brittle when they encounter actual interfaces or disorder on a chip.

Mira: It shifts the focus from just material properties to the geometry of how those materials are arranged and terminated, making it easier to design training sets that reflect device relevant complexity.

Kai: Ultimately, this framework aims to make the creation of realistic 2D and interface datasets systematic and traceable, which should improve how we use AI in materials science across the board <ref:2603.27886#pg1>.

Vsevolod Biryukov, Kamal Choudhary, Timur Bazhirov

Exabyte Inc. · Department of Materials Science and Engineering, Johns Hopkins University · Department of Electrical and Computer Engineering, Johns Hopkins University

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

Submitted: 2026-03-29

Updated: 2026-03-29

Comments: 23 pages, 7 figures, 1 table

DOI: 10.1038/s41598-026-72825-z

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals…

Key concepts

Mat3ra-2D
An open-source framework for designing realistic 2D materials and interfaces. It allows users to systematically construct complex structures like slabs, surfaces, and interfaces by following defined workflows rather than just selecting bulk data.
Provenance
A record that tracks every choice made during structure generation, such as orientation, termination type, layer count, and strain. This detailed history ensures that any generated structure can be traced back to its exact creation steps for transparency.
Structure Generation Workflows
Systematic methods for building realistic structures. These include specific procedures for creating orientation-specific slabs or assembling film/substrate interfaces, allowing users to control parameters like strain and registry explicitly.

Terminology

Summary

Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate <ref:2603. p. 0

Where the paper says it

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity.

27886#pg2> The paper presents Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity <ref:2603. p. 0
Where the paper says it

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity.

27886#pg2>.

Framework Architecture

Mat3ra-2D is organized as an implementation stack for practical generation of realistic structures, such as slabs, surfaces, interfaces, and defective configurations, with explicit provenance <ref:2603.27886#pg2> Rather than a single monolithic codebase, the methodology is distributed across standards, common data, implementation layers, and usage examples <ref:2603. p. 1

Where the paper says it

1 Overview Mat3ra-2D is organized as an implementation stack for practical generation of realistic structures, such as slabs, surfaces, interfaces, and defective configurations, with explicit provenance. Rather than presenting a single monolithic codebase, the methodology is distributed across standards, common data, implementation layers, and usage examples, each with a distinct role in making realistic structure generation reusable and reproducible.

27886#pg2>. The components include:

  1. Standards package mat3ra-esse [24, 25], which provides the schemas, ontology, categorization, and exchange format used throughout the ecosystem <ref:2603. p. 2

    Where the paper says it

    2 Components 2. 1 Standards The lowest layer of the stack is the standards package mat3ra-esse [24, 25], which provides the schemas, ontology, categorization, and exchange format used throughout the ecosystem.

    27886#pg2>

  2. Commonly-used data mat3ra-standata [26], which distributes commonly used example materials and related records that adhere to the shared standards <ref:2603. p. 2

    Where the paper says it

    In the present work, we use that foundation as the substrate for realistic 2D structure generation rather than reintroducing the full categorization itself. 2 Commonly-used data The next layer is mat3ra-standata [26], which distributes commonly used example materials and related records that adhere to the shared standards.

    27886#pg2>

  3. OOD abstraction layer mat3ra-code, which provides implementations of the abstract level of the standardized entities as Python and JavaScript/TypeScript classes <ref:2603. p. 2

    Where the paper says it

    In practice, this package serves as a reusable source of input structures for workflows and notebooks, allowing examples to begin from consistent reference materials rather than redefining them locally. 3 OOD abstraction layer Above the standards layer, mat3ra-code provides implementations of the abstract level of the standardized entities as Python and JavaScript/TypeScript classes.

    27886#pg2>

  4. Materials design implementation mat3ra-made [27], which contains the concrete materialsdesign functionality used to construct realistic structures <ref:2603. p. 2

    Where the paper says it

    4 Materials design implementation The main implementation layer is mat3ra-made [27]. This package contains the concrete materialsdesign functionality used to construct realistic structures: material objects, builders, analyzers, and modifiers for operations such as slab generation, interface construction, strain handling, and defect creation.

    27886#pg2>

  5. Notebook examples and browser access mat3ra-api-examples [28], which shows how the stack is used in practice through reusable Jupyter notebooks <ref:2603. p. 2

    Where the paper says it

    5 Notebook examples and browser access The user-facing examples are distributed through mat3ra-api-examples [28]. This layer shows how the stack is used in practice through reusable Jupyter notebooks.

    27886#pg2>

Structure Generation Workflows

The framework enables systematic generation of structures that are central to applications but underrepresented in bulk-centric datasets <ref:2603.27886#pg2>. Key workflows demonstrated include:

** Constructing orientation-specific slabs with explicit termination control is straightforward, for example, a SrTiO3(110) slab is created by importing bulk SrTiO3 from the reference data repository, then applying a slab builder with Miller indices (1,1,0), specifying the number of layers and desired termination formula <ref:2603.27886#pg2>.**

Assembly of film/substrate interfaces without strain matching follows a define–build workflow where slabs are defined independently and then combined with a target interfacial gap and optional lateral shift <ref:2603. p. 4

Where the paper says it

2 Example 2: constructing interfaces without strain matching Assembly of film/substrate interfaces without strain matching follows a define–build workflow. For example, a Ge(001)/Si(001) interface is constructed by defining both slabs independently (each with Miller indices (001), thickness, and vacuum), then combining them with a target interfacial gap (e.

27886#pg2>.

For interfaces requiring commensurate matching, the framework provides a define–refine–build workflow involving defining film and substrate slabs, refining by enumerating commensurate matches using a ZSL analyzer, and then building the selected configuration with recorded metadata <ref:2603. p. 7

Where the paper says it

(a) Graphene monolayer (b) Ni(001) slab (c) Strain–size trade-off (d) Selected Gr/Ni(001) interface Figure 4: Graphene/Ni(001) interface construction workflow: (a) Starting graphene monolayer film, (b) Ni(001) slab substrate, (c) ZSL analyzer output showing strain–size trade-off for candidate configurations (each point represents a commensurate match ranked by strain percentage and interface area), (d) Selected interface structure with commensurate matching applied. The workflow follows define–refine–build: define film and substrate slabs, refine by enumerating and ranking matches, then build the selected configuration with recorded metadata.

27886#pg2>.

Reproducibility and Accessibility

The paper emphasizes the combination of provenance-aware generation with openly accessible execution environments to ensure transparency <ref:2603.27886#pg2>.

  1. Provenance records capture the sequence of choices used to create a structure, including orientation, termination, thickness, strain, registry, and intermediate ranking criteria <ref:2603. p. 1

    Where the paper says it

    In such cases, the relevant scientific degrees of freedom often arise from how a structure is created rather than from stoichiometry alone: orientation, termination, layer count, vacuum spacing, strain state, stacking registry, defect type, and post-processing steps can all alter the resulting material model. As a result, generating realistic structures for computation is not simply a matter of selecting a material from a database, but of constructing and documenting a sequence of transformations that leads to the target geometry.

    27886#pg2>

  2. Generic notebooks are designed to be educational and adaptable by accepting multiple input parameters and providing functionality for loading materials from the reference data repository <ref:2603. p. 8

    Where the paper says it

    , slab creation, interface construction, passivation). These notebooks are designed to be educational and adaptable: they include functionality for loading materials from the reference data repository, previewing structures, saving results, and provide default settings and parameters.

    27886#pg2>

  3. The notebook layer is designed to operate in conventional Python environments and in browser-based runtimes, such as JupyterLite using Pyodide, enabling inspection and execution without local installation <ref:2603. p. 3

    Where the paper says it

    The notebook layer is designed to operate in conventional Python environments and in browser-based runtimes. In particular, the same examples can be executed through JupyterLite using Pyodide, enabling inspection and execution without local installation.

    27886#pg2>.

Implications for AI/ML Data

Mat3ra-2D supports the creation of training and benchmark datasets that better reflect the target deployment domain <ref:2603.27886#pg2>. The framework helps address the limitation where many machine-learning models are dominated by idealized three-dimensional crystals, while real-world performance depends on surfaces, reconstructions, disorder, and defects <ref:2603. p. 0

Where the paper says it

At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13]. A persistent limitation is that many machine-learning models and benchmark datasets are dominated by idealized three-dimensional crystals, while real-world performance often depends on surfaces, reconstructions, disorder, and defects, as well as heterogeneous interfaces that determine transport, catalysis, and stability [14, 15].

27886#pg2>. By making structure generation explicit and reproducible with provenance, datasets can be refined in a targeted way, and training or benchmark collections can be regenerated as assumptions evolve <ref:2603. p. 1
Where the paper says it

1 Overview Mat3ra-2D is organized as an implementation stack for practical generation of realistic structures, such as slabs, surfaces, interfaces, and defective configurations, with explicit provenance. Rather than presenting a single monolithic codebase, the methodology is distributed across standards, common data, implementation layers, and usage examples, each with a distinct role in making realistic structure generation reusable and reproducible.

27886#pg2>.

Practical Considerations

The framework clarifies the scope by providing machinery to define, build, annotate, and share realistic structures <ref:2603.27886#pg2>. However, it does not remove the scientific judgment required to decide which structures are representative or which fidelity level is appropriate for downstream simulation <ref:2603. p. 2

Where the paper says it

Its role is to provide a shared machine-readable representation for structures, inputs, and provenance so that higher-level tools operate on consistent concepts. These standards, together with representative schema and JSON examples, are based on the M-CODE categorization and ontology described in the corresponding manuscript [21].

27886#pg2>. Reproducible generation is not the same as validated generation, as a structure can still be a poor approximation if the wrong termination or defect model was chosen <ref:2603. p. 1
Where the paper says it

This challenge is particularly acute for 2D and surface-dominated systems. In such cases, the relevant scientific degrees of freedom often arise from how a structure is created rather than from stoichiometry alone: orientation, termination, layer count, vacuum spacing, strain state, stacking registry, defect type, and post-processing steps can all alter the resulting material model.

27886#pg2>. The framework is intended to complement existing computational materials infrastructure rather than replace it <ref:2603. p. 1
Where the paper says it

Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling. The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings.

27886#pg2>.

Future Directions

Future work should focus on validation, scalable dataset design, and broader access through open notebooks <ref:2603.27886#pg2>. One direction is to systematically compare generated interfaces against experimental observations to identify reliable classes of approximations <ref:2603.27886#pg2>. Another is benchmark suite development rather than brute-force coverage, as realistic structure spaces are too large to enumerate exhaustively <ref:2603.27886#pg2>. The long-term value depends on combining accessibility with scientific rigor to make realistic structure generation both easier to use and easier to trust <ref:2603.27886#pg2>.

Conclusion

This work presented Mat3ra-2D as a framework for constructing realistic 2D materials and related structures with explicit, machine-readable provenance <ref:2603.27886#pg2>. By prioritizing realism, provenance, accessibility, and reuse, Mat3ra-2D aims to enable dataset design and modeling workflows that better reflect device-relevant materials complexity and improve the practical utility of AI/ML models <ref:2603. p. 1

Where the paper says it

Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling. The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings.

27886#pg2>.

Data Availability

M-CODE and the associated schemas are maintained as open-source data standards <ref:2603.27886#pg2>. The reference implementation is accessible as a package on GitHub and distributed via PyPI as mat3ra-esse <ref:2603. p. 2

Where the paper says it

2 Components 2. 1 Standards The lowest layer of the stack is the standards package mat3ra-esse [24, 25], which provides the schemas, ontology, categorization, and exchange format used throughout the ecosystem.

27886#pg2>. Related ecosystem packages and notebooks are available on GitHub and PyPI, including mat3ra-standata, mat3ra-made, and mat3ra-api-examples <ref:2603. p. 2
Where the paper says it

2 Commonly-used data The next layer is mat3ra-standata [26], which distributes commonly used example materials and related records that adhere to the shared standards. In practice, this package serves as a reusable source of input structures for workflows and notebooks, allowing examples to begin from consistent reference materials rather than redefining them locally.

27886#pg2>.

References

[1] Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder <ref:2603.27886#pg2>

[2] Stefano Curtarolo et al. Aflowlib. org: A distributed materials properties repository from high-throughput ab initio calculations <ref:2603. p. 0

Where the paper says it

Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7]. At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13].

27886#pg2>

[3] James E Saal et al. Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd) <ref:2603. p. 0

Where the paper says it

1 Introduction Modern materials research increasingly relies on digital data and automation to shorten the time from hypothesis to validated insight. Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7].

27886#pg2>

[4] Giovanni Pizzi et al. Aiida: automated interactive infrastructure and database for computational science <ref:2603. p. 0

Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

27886#pg2>

[5] NOMAD laboratory: A European Centre of Excellence <ref:2603.27886#pg2>

[6] Kamal Choudhary et al. The joint automated repository for various integrated simulations (jarvis) for data-driven materials design <ref:2603. p. 1

Where the paper says it

Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling. The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings.

27886#pg2>

[7] Jaehyung Lee et al. Agapi-agents: An open-access agentic ai platform for accelerated materials design on atomgpt <ref:2603. p. 1

Where the paper says it

Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling. The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings.

27886#pg2>

[8] Sten Haastrup et al. The computational 2d materials database: High-throughput modeling and discovery of atomically thin crystals <ref:2603. p. 0

Where the paper says it

Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7]. At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13].

27886#pg2>

[9] Filip A Rasmussen and Kristian S Thygesen. Computational 2d materials database: electronic structure of transition-metal dichalcogenides and oxides <ref:2603. p. 1

Where the paper says it

In such cases, the relevant scientific degrees of freedom often arise from how a structure is created rather than from stoichiometry alone: orientation, termination, layer count, vacuum spacing, strain state, stacking registry, defect type, and post-processing steps can all alter the resulting material model. As a result, generating realistic structures for computation is not simply a matter of selecting a material from a database, but of constructing and documenting a sequence of transformations that leads to the target geometry.

27886#pg2>

[10] Kiran Mathew et al. Mpinterfaces: A materials project based python tool for high-throughput computational screening of interfacial systems <ref:2603. p. 0

Where the paper says it

1 Introduction Modern materials research increasingly relies on digital data and automation to shorten the time from hypothesis to validated insight. Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7].

27886#pg2>

[11] Kamal Choudhary et al. High-throughput identification and characterization of two-dimensional materials using density functional theory <ref:2603. p. 0

Where the paper says it

1 Introduction Modern materials research increasingly relies on digital data and automation to shorten the time from hypothesis to validated insight. Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7].

27886#pg2>

[12] Kamal Choudhary et al. Efficient computational design of two-dimensional van der waals heterostructures: Band alignment, lattice mismatch, and machine learning <ref:2603. p. 0

Where the paper says it

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity.

27886#pg2>

[13] Kamal Choudhary and Kevin F Garrity. Intermat: accelerating band offset prediction in semiconductor interfaces with dft and deep learning <ref:2603. p. 1

Where the paper says it

As a result, generating realistic structures for computation is not simply a matter of selecting a material from a database, but of constructing and documenting a sequence of transformations that leads to the target geometry. As demonstrated in the recent study[16], comparing 16 graphbased machine-learned force fields, the lack of data on structural features relevant to real-world semiconductor applications can hinder the model’s applicability to interfaces.

27886#pg2>

[14] Logan Ward et al. A generalpurpose machine learning framework for predicting properties of inorganic materials <ref:2603. p. 0

Where the paper says it

At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13]. A persistent limitation is that many machine-learning models and benchmark datasets are dominated by idealized three-dimensional crystals, while real-world performance often depends on surfaces, reconstructions, disorder, and defects, as well as heterogeneous interfaces that determine transport, catalysis, and stability [14, 15].

27886#pg2>

[15] Olexandr Isayev et al. Universal fragment descriptors for predicting properties of inorganic crystals <ref:2603. p. 0

Where the paper says it

Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7]. At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13].

27886#pg2>

[16] Daniel Wines and Kamal Choudhary. Chips-ff: Evaluating universal machine learning force fields for material properties <ref:2603. p. 1

Where the paper says it

As a result, generating realistic structures for computation is not simply a matter of selecting a material from a database, but of constructing and documenting a sequence of transformations that leads to the target geometry. As demonstrated in the recent study[16], comparing 16 graphbased machine-learned force fields, the lack of data on structural features relevant to real-world semiconductor applications can hinder the model’s applicability to interfaces.

27886#pg2>

[17] Protik Das et al. Accessible computational materials design with high fidelity and high throughput <ref:2603. p. 0

Where the paper says it

1 Introduction Modern materials research increasingly relies on digital data and automation to shorten the time from hypothesis to validated insight. Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7].

27886#pg2>

[18] T. Bazhirov. Data-centric online ecosystem for digital materials science <ref:2603. p. 1

Where the paper says it

The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings. In this sense, it builds on our prior work on accessible computational materials design [17], data-centric ecosystem design [18], practical categorization of computational models [19], interpretable materials ML [20], and, most directly, M-CODE as a categorization and ontology layer for realistic structures [21].

27886#pg2>

[19] A. Zech and T. Bazhirov. Catecom: A practical data-centric approach to categorization of computational models <ref:2603. p. 1

Where the paper says it

The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings. In this sense, it builds on our prior work on accessible computational materials design [17], data-centric ecosystem design [18], practical categorization of computational models [19], interpretable materials ML [20], and, most directly, M-CODE as a categorization and ontology layer for realistic structures [21].

27886#pg2>

[20] J. Dean et al. Interpretable machine learning for materials design <ref:2603. p. 0

Where the paper says it

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity.

27886#pg2>

[21] Vsevolod Biryukov, Kamal Choudhary, and Timur Bazhirov. M-code: Materials categorization via ontology, dimensionality and evolution <ref:2603. p. 1

Where the paper says it

The framework combines shared data standards, reusable transformation logic, provenance-aware workflows, and example-driven access paths so that realistic structures can be generated, inspected, reused, and adapted across different computational settings. In this sense, it builds on our prior work on accessible computational materials design [17], data-centric ecosystem design [18], practical categorization of computational models [19], interpretable materials ML [20], and, most directly, M-CODE as a categorization and ontology layer for realistic structures [21].

27886#pg2>

[22] Shyue Ping Ong et al. Python materials genomics (pymatgen): A robust, open-source python library for materials analysis <ref:2603. p. 1

Where the paper says it

com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks. The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE [22, 23], while also supporting browser-based execution for lower-friction access to representative workflows.

27886#pg2>

[23] Ask Hjorth Larsen et al. The atomic simulation environment—a python library for working with atoms <ref:2603.27886#pg2>

[24] mat3ra-esse. Mat3ra.com, 2026 <ref:2603. p. 0

Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

p. 0
Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

p. 0
Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

p. 0
Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

p. 0
Where the paper says it

com), Walnut Creek, CA 94596, USA †Department of Materials Science and Engineering, Johns Hopkins University, Baltimore, MD 21218, USA ‡Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218, USA §Corresponding author: timur@mat3ra. com 1 arXiv:2603.

27886#pg2>

[25] @mat3ra/esse. Mat3ra.com, 2026 <ref:2603.27886#pg2>

[26] mat3ra-standata. Mat3ra.com, 2026 <ref:2603.27886#pg2>

[27] mat3ra-made. Mat3ra.com, 2026 <ref:2603.27886#pg2>

[28] mat3ra-api-examples. Mat3ra.com, 2026 <ref:2603.27886#pg2>

--- Page 1 ---

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate <ref:2603.27886#pg2> We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity <ref:2603. p. 0

Where the paper says it

AI-ready design of realistic 2D materials and interfaces with Mat3ra-2D Vsevolod Biryukov∗ Kamal Choudhary †‡ Timur Bazhirov∗§ Abstract Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defectdriven complexity.

27886#pg2>.

--- Page 2 ---

This challenge is particularly acute for 2D and surface-dominated systems <ref:2603.27886#pg2>. In such cases, the relevant scientific degrees of freedom often arise from how a structure is created rather than from stoichiometry alone: orientation, termination, layer count, vacuum spacing, strain state, stacking registry, defect type <ref:2603. p. 1

Where the paper says it

This challenge is particularly acute for 2D and surface-dominated systems. In such cases, the relevant scientific degrees of freedom often arise from how a structure is created rather than from stoichiometry alone: orientation, termination, layer count, vacuum spacing, strain state, stacking registry, defect type, and post-processing steps can all alter the resulting material model.

27886#pg2>.

--- Page 3 ---

As a result, generating realistic structures for computation is not simply a matter of selecting a material from a database, but of constructing and documenting a sequence of transformations that leads to the target geometry <ref:2603.27886#pg2>. Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling <ref:2603. p. 1

Where the paper says it

As demonstrated in the recent study[16], comparing 16 graphbased machine-learned force fields, the lack of data on structural features relevant to real-world semiconductor applications can hinder the model’s applicability to interfaces. Mat3ra-2D addresses these needs with an open-source framework for practical design of realistic 2D materials and related structures, including slabs, interfaces, defects, and other lowdimensional motifs relevant to device modeling.

27886#pg2>.

--- Page 4 ---

The approach combines: (1) well-defined standards for storing and exchanging materials data with a modular implementation of core concepts and (2) transformation workflows expressed as configuration-builder pipelines that preserve provenance and metadata <ref:2603.27886#pg2>

--- Page 5 ---

We implement typical structure generation tasks, such as constructing orientationspecific slabs or strain-matching interfaces, in reusable Jupyter notebooks that serve as both interactive documentation and templates for reproducible runs <ref:2603.27886#pg2>.

--- Page 6 ---

Mat3ra-2D is part of Mat3ra.com, a collaborative platform for computational materials R&D <ref:2603. p. 1

Where the paper says it

Mat3ra-2D is part of Mat3ra. com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks.

p. 1
Where the paper says it

Mat3ra-2D is part of Mat3ra. com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks.

27886#pg2> and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks <ref:2603. p. 1
Where the paper says it

Mat3ra-2D is part of Mat3ra. com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks.

p. 1
Where the paper says it

Mat3ra-2D is part of Mat3ra. com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks.

27886#pg2>.

--- Page 7 ---

The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE <ref:2603. p. 1

Where the paper says it

com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks. The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE [22, 23], while also supporting browser-based execution for lower-friction access to representative workflows.

27886#pg2> while also supporting browser-based execution for lower-friction access to representative workflows <ref:2603. p. 1
Where the paper says it

com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks. The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE [22, 23], while also supporting browser-based execution for lower-friction access to representative workflows.

p. 1
Where the paper says it

com, a collaborative platform for computational materials R&D [18], and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks. The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE [22, 23], while also supporting browser-based execution for lower-friction access to representative workflows.

27886#pg2>.

--- Page 8 ---

The resulting focus on explicit structure-generation choices is aligned with our broader effort to make materials datasets and models more traceable and reusable across the digital materials workflow <ref:2603.27886#pg2>.

--- Page 9 ---

Mat3ra-2D is part of Mat3ra.com, a collaborative platform for computational materials R&D <ref:2603.27886#pg2> and is focused on organizing realistic structure generation as a reusable software stack spanning standards, common data, implementation packages, and interactive notebooks <ref:2603.27886#pg2>.

--- Page 10 ---

The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE <ref:2603.27886#pg2> while also supporting browser-based execution for lower-friction access to representative workflows <ref:2603.27886#pg2>.

--- Page 11 ---

The resulting focus on explicit structure-generation choices is aligned with our broader effort to make materials datasets and models more traceable and reusable across the digital materials workflow <ref:2603.27886#pg2>.

--- Page 12 ---

Mat3ra-2D enables systematic creation and organization of realistic 2D- and interface-aware datasets for AI/ML-ready applications <ref:2603.27886#pg2>.

--- Page 13 ---

The remainder of this manuscript is organized as follows. Section 2 outlines the overall approach and the main components of the ecosystem <ref:2603. p. 1

Where the paper says it

The remainder of this manuscript is organized as follows. Section 2 outlines the overall approach and the main components of the ecosystem.

p. 1
Where the paper says it

The remainder of this manuscript is organized as follows. Section 2 outlines the overall approach and the main components of the ecosystem.

27886#pg2>.

--- Page 14 ---

Section 3 summarizes representative outcomes and example workflows enabled by Mat3ra-2D <ref:2603.27886#pg2>.

--- Page 15 ---

The remainder of this manuscript is organized as follows. Section 2 outlines the overall approach and the main components of the ecosystem <ref:2603.27886#pg2>.

--- Page 16 ---

Figure 4: Graphene/Ni(001) interface construction workflow: (a) Starting graphene monolayer, (b) Ni(001) slab substrate, (c) ZSL analyzer output showing strain–size trade-off for candidate configurations (each point represents a commensurate match ranked by strain percentage and interface area), (d) Selected Gr/Ni(001) interface structure with commensurate matching applied <ref:2603.27886#pg2>

--- Page 17 ---

The same exact workflows can be executed in standard Python environments (JupyterLab, IPython) or directly in a web browser via Pyodide, lowering the setup and getting-started barriers while remaining interoperable with established toolkits such as pymatgen and ASE <ref:2603.27886#pg2>.

--- Page 18 ---

Mat3ra-2D is motivated by a practical constraint in data-driven materials modeling: the structures most relevant to real devices and processes are rarely ideal bulk crystals <ref:2603.27886#pg2>.

--- Page 19 ---

A central design choice in Mat3ra-2D is to treat realistic structure generation as a composition of reusable workflow steps rather than as a collection of isolated scripts <ref:2603.27886#pg2>. The approach is designed to interoperate with established Python tooling for materials analysis and atomistic modeling, including pymatgen and ASE <ref:2603.27886#pg2>.

Pruning redundant reasoning tokens

The paper presents Mat3ra-2D as a framework for constructing realistic 2D materials and related structures with explicit, machine-readable provenance <ref:2603.27886#pg2> It is motivated by the need to capture structures most relevant to experiments and devices, including monolayers, slabs… p. 0

Where the paper says it

Large-scale open databases and workflow management systems have enabled high-throughput computational screening and accelerated discovery across broad application domains [1–7]. At the same time, the processes that govern materials properties in practice combined with the rapid growth of atomically thin materials has created a need for datasets and tooling that go beyond ideal bulk crystals and capture the structures most relevant to experiments and devices, including monolayers, slabs, interfaces and heterostructures with defects and disorder[8–13].

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. Data-driven design of device-relevant materials will be enabled by Mat3ra-2D to create training and benchmark datasets that better reflect the target deployment domain. This allows AI/ML models to be trained on realistic structures including surfaces, reconstructions, disorder, and defects rather than just ideal bulk crystals.

  2. Automated structure generation workflows can be implemented in any browser via JupyterLite using Pyodide. This lowers the barrier to adoption by enabling users to inspect and execute without local installation, allowing AI researchers to rapidly test structure generation parameters directly within a web application context.

  3. AI/ML models will gain improved transferability across different material systems by incorporating explicit provenance into structure creation. When structures carry explicit provenance, they become inspectable and rerunnable rather than opaque, which helps in diagnosing why models fail on real-world interfaces versus ideal crystals.

  4. The system can systematically explore the combinatorial design space of 2D materials, such as strain-matched interfaces, by utilizing a define–refine–build workflow. This allows for the systematic generation of realistic structures through tools like the ZSL interface analyzer to find optimal configurations based on trade-offs between strain minimization and computational cost.

  5. AI systems can perform inverse design tasks guided by structure generation workflows. By having structures with explicit provenance, researchers can refine datasets in a targeted way, allowing them to create new training collections as assumptions evolve based on the insights gained from the generated data.

Abstract

Artificial intelligence (AI) and machine learning (ML) models in materials science are predominantly trained on ideal bulk crystals, limiting their transferability to real-world applications where surfaces, interfaces, and defects dominate. We present Mat3ra-2D, an open-source framework for the rapid design of realistic two-dimensional materials and related structures, including slabs and heterogeneous interfaces, with support for disorder and defect-driven complexity. The approach combines: (1) well-defined standards for storing and exchanging materials data with a modular implementation of core concepts and (2) transformation workflows expressed as configuration-builder pipelines that preserve provenance and metadata. We implement typical structure generation tasks, such as constructing orientation-specific slabs or strain-matching interfaces, in reusable Jupyter notebooks that serve as both interactive documentation and templates for reproducible runs. To lower the barrier to adoption, we design the examples to run in any web browser and demonstrate how to incorporate these developments into a web application. Mat3ra-2D enables systematic creation and organization of realistic 2D- and interface-aware datasets for AI/ML-ready applications.

Sources

Related papers