[Paper Review] M-CODE: Materials Categorization via Ontology, Dimensionality and Evolution

Vsevolod Biryukov, Kamal Choudhary+1 more|arXiv (Cornell University)|Feb 16, 2026

Machine Learning in Materials Science0 citations

TL;DR

M-CODE presents a compact, ontology-driven categorization system for realistic materials structures, linking domain concepts to reusable software entities with provenance-aware build transformations, implemented as open JSON schemas and code.

ABSTRACT

The rapid advancement of artificial intelligence in materials science requires data standards and data management practices that can capture the complexity of real-world structures, including surfaces, interfaces, defects, and dimensionality reduction. We present M-CODE - Materials Categorization via Ontology, Dimensionality and Evolution - a compact categorization system that links materials-science-specific terminology to a set of reusable concepts as building blocks and provenance-aware transformations. M-CODE classifies structures by dimensionality, structural complexity (from pristine to compound pristine, defective, and processed), and variants that capture common structure creation and evolution approaches. A practical implementation of the categorization is provided in an open-source codebase that includes JSON schemas, examples, and Python and TypeScript types/interfaces, designed to support reproducible dataset generation, validation, and community contributions.

Motivation & Objective

• Bridge the gap between idealized training structures and realistic, defective, and interface-rich materials.
• Define a compact, domain-grounded categorization of structures by dimensionality, domain, and variants.
• Provide an ontology of entities and operations that can be mapped to JSON schemas and software interfaces.
• Capture provenance of structure construction to support reproducible dataset generation and sharing.

Proposed method

• Develop a four-category entity/operation ontology to describe building blocks and transformations for materials structures.
• Use JSON Schemas to validate and exchange configurations; auto-generate Python and TypeScript bindings from schemas.
• Represent materials generation as configurations, builders, and results with explicit provenance metadata.
• Define evolution pathways (pristine, compound pristine, defective, processed) to connect structure targets to reproducible build workflows.
• Provide a reference implementation and open schemas (mat3ra-esse) for validation and reuse.

Figure 2: Representative target structure classes grouped by domain and annotated with the corresponding M-CODE tags. See Table 5 . Pristine Structures (top row): MoS 2 monolayer, SrTiO 3 slab, Si nanowire, and graphene nanoribbon. Compound Pristine Structures (second row): Si/SiO 2 /HfO 2 /TiN hete

Experimental results

Research questions

• RQ1 How can realistic, low-dimensional, and heterogeneous material structures be standardized for dataset generation and workflow reproducibility?
• RQ2 Can a compact, ontology-based categorization plus provenance-aware transformations improve interoperability across tools and databases in materials data frameworks?
• RQ3 How can domain terminology be mapped to reusable software entities to enable automated workflow routing and FAIR data practices?

Key findings

• A compact categorization framework classifies structures by domain, dimensionality, category, and variants with stable M-CODE tags.
• An open, software-oriented ontology of entities and operations can be mapped to JSON schemas and language bindings for reproducible builds.
• Provenance-aware configurations and transformations enable explicit regeneration of structures with permissive metadata for extra context.
• The approach supports category-aware workflow routing and FAIR data principles, enabling AI/ML readiness and reproducible datasets.
• The work provides canonical schema artifacts and a reference implementation distributed as mat3ra-esse for validation and reuse.