tm-spec

Lit-Review: Metalanguages for Scientific Data Description

Date: 2026-05-09. Audience: self-study prior to extending TM-Spec. Goal: understand what is already standardised, to avoid reinventing the wheel and instead borrow best practices layer by layer.

---

TL;DR

There is no universal “physics/chemistry language.” What exists is a stack of established domain languages, organised by layer. The most “code-like” ones in our zone:

Domain	Best language	Why
Reaction dynamics (CRN/kinetics)	Antimony 3	Text like J0: A->B; k1*A, converts to SBML, reads like code
Atomistics (DFT/MD inputs)	ASE Python	Python as de facto DSL over QE/VASP/GPAW/CP2K
Defect chemistry	Kröger-Vink	Defect equations with charges and sites
Hydrothermal environment	PHREEQC input	KEYWORD-blocks + balanced equations, all in one file
Phase thermodynamics	CALPHAD TDB	A real DSL for Compound Energy Formalism
Crystal structure	AFLOW prototype + CIF	AB_tP4_129_a_c — compact signature
Cross-DFT interoperability	NOMAD Metainfo	Parsers for ~40 codes
Materials database queries	OPTIMADE	Unified filter language
Reproducibility	AiiDA	Workflow + provenance graph

---

Part 1. Generic / Organic Languages

1.1. Structure (static)

• SMILES / SMARTS / SMIRKS — linear string for molecules and reactions (CC(=O)O = acetic acid). De facto standard in organic chemistry.
• InChI / InChIKey — canonical molecular name (hash-like).
• CIF (IUCr) — crystallography; heavy but universal.
• extxyz / POSCAR — atomistics; minimalist.
• CML (Chemical Markup Language) — XML-based; has lost popularity.

1.2. Reaction networks — “programming like code”

Antimony — the most readable

A text front-end for SBML, developed by the Sauro lab (U. Washington):

J0: A -> B; k1*A
J1: B -> C; k2*B
k1 = 0.1; k2 = 0.5; A = 10

Antimony 3 (2024–2025) added FBC, Layout, annotations, and distributions. The closest to a “reads like a program” answer for CRN/dynamics.

Rule-based

• BNGL (BioNetGen Language, Faeder/Hlavacek) — rule-based for combinatorial spaces (signalling pathways, protein states).
• Kappa — a close analogue of BNGL, from a separate group (Danos/Krivine).
• Both are powerful but unreadable at scale. Antimony has a modular structure; Kappa does not, making large Kappa models hard to maintain.

Kinetics

• SBML — XML under the hood, not meant for humans; Antimony is its text front-end.
• Cantera YAML (≥2.5) — gas-phase/catalytic kinetics, readable: reactions, thermodynamics, and transport in a single file.
• MobsPy (PLOS Comp Bio 2025) — Python DSL for biochemical reaction networks, object-oriented on top of the BNGL idea.

1.3. DFT / atomistics — no unified language

Each code has its own format: QE namelists, VASP INCAR, ABACUS INPUT, GPAW Python, CP2K input sections, ORCA keyword lines. Unification attempts:

• NOMAD Metainfo (FAIRmat NFDI) — common schema with parsers for ~40 codes (QE, VASP, GPAW, CP2K, ABINIT, FHI-aims, ORCA, Gaussian, LAMMPS, GROMACS…). This is a schema + parsers, not a language you write input in — it is what your input is converted into for interoperability.
• OPTIMADE — query language for materials databases (elements HAS "Fe" AND nelements=2). Describes data, not calculations.
• QCSchema (MolSSI) — JSON schema for QM calculations (input/output/wfn). Used in QCArchive, Psi4, and the MolSSI stack.
• ASE (Atomic Simulation Environment) — Python as de facto DSL: Atoms(), calc=Espresso(...), MDLogger, NEB. A single script reads identically regardless of backend (QE/VASP/GPAW). In our project ASE is the primary “metalanguage.”
• AiiDA — workflow + provenance graph (every input/output node hashed, graph is reproducible). A reproducibility metalanguage, not a process description language.
• atomate2 / Fireworks — Python workflows on top of pymatgen.

1.4. Force fields

• SMIRNOFF (Open Force Field Initiative) — XML, explicitly readable: torsions described by SMARTS pattern + parameters. One file = one force field.

1.5. Experiment / laboratory

• ISA-Tab / ISA-JSON — investigation/study/assay structure.
• ELN formats (eLabFTW, RSpace) — schema-friendly notebooks.
• NeXus / HDF5 — for neutron/synchrotron data.

1.6. Recent developments 2025–2026

1. Antimony 3 — major update; now covers constraint-based modelling (FBA), distributions, and layout. Closest to a “program in a language” for chemical dynamics.
2. MobsPy — Python DSL for biochemical reaction networks, object-oriented on top of the BNGL idea.
3. LangSim (MPIE) — LLM front-end over atomistic simulation: write in plain English, an agent calls ASE/pyiron underneath. Not a metalanguage but a natural-language interpreter into ASE.
4. Foundation models for atomistic simulation (Nat Rev Chem 2025) — a review showing that the field is not converging on a unified language, but on unified representations (graph + species embeddings).

---

Part 2. Inorganic / Materials Science

2.1. Crystal structure

• ⭐ CIF (IUCr standard) — text, but not directly human-friendly: space group, Wyckoff positions, ADP. Readable with practice.
• magCIF — CIF extension for magnetic structures (modulation vector, magnetic SG).
• AFLOW prototype labels — compact signature: A2B_cF24_225_a_c = composition + Pearson symbol + space group + Wyckoff letters. Closest to a “filename that tells you everything.”
• Wyckoff / Pearson / Strukturbericht (B1, B2, C1…) — old but still alive in the literature.
• POSCAR / extxyz — for calculations; minimalist.

2.2. Defect chemistry — ⭐ a genuine DSL

Kröger-Vink notation — equations that function as code:

nil → V_M''  + V_X^••           (Schottky)
M_M^× + X_X^× → V_M''  + V_X^•• + M_i^•• + X_i''   (Frenkel)
2·Fe_Fe^× + ½ O₂(g) → 2·Fe_Fe^•  + V_Fe''  + O_O^×   (acceptor doping)

Written as a reaction equation using standard notation:

• superscript = charge relative to the ideal lattice (• positive, ‘ negative, × neutral),
• subscript = site.

90% readable after one page of a textbook. No official ASCII standard exists, but pymatgen.analysis.defects and doped (Walsh group, 2024) encode this in Python objects.

2.3. Surfaces

• Wood notation: (2×2)R45°, c(4×2) — for commensurate reconstructions.
• Matrix notation — for incommensurate: $\binom{2 \, 1}{-1 \, 2}$.
• ASE slab API — Python DSL: surface(bulk, indices=(1,1,1), layers=4, vacuum=15) reads like a function call.

2.4. Magnetic structure

• magCIF + MAGNDATA (Bilbao Crystallographic Server) — a genuine DSL for magnetic structures: propagation vector k=(0,0,1/2), irreducible representations, magnetic space groups (Belov-Neronova-Smirnova).
• Mantid scripting — for neutron data.

2.5. Phase thermodynamics — ⭐ a genuine DSL

CALPHAD TDB format (Thermo-Calc, OpenCalphad, pycalphad, ESPEI) — text-based databases:

PHASE LIQUID:L %  1  1.0 !
CONSTITUENT LIQUID:L : FE,S : !
PARAMETER G(LIQUID,FE;0)  298.15  GFEL;  6000 N !
PARAMETER L(LIQUID,FE,S;0)  298.15  -104888.7+0.338*T;  6000 N !

Compound Energy Formalism — a real program in a DSL for thermodynamic modelling. Recent (2025): coupling with MLIPs (MACE-MP / UMA) for automatic parameter fitting. For our Fe-S, Ni-S context, the CALPHAD databases SGTE PURE5 + SSUB already contain FeS, FeS₂, Fe₃S₄.

2.6. Aqueous / hydrothermal systems — ⭐⭐ relevant to our project

PHREEQC input (USGS) — already installed (phreeqpython). A pure KEYWORD-DSL with balanced chemical equations:

SOLUTION 1
    pH 6.5
    temp 90
    Fe(2) 1.0e-3
    S(-2) 5.0e-4
SOLUTION_SPECIES
    Fe+2 + HS- = FeHS+ ; log_k 5.30
PHASES
    Mackinawite
    FeS + H+ = Fe+2 + HS- ; log_k -3.6
KINETICS
    Pyrite_oxidation
    -m0  1e-6
    -formula  FeS2 + 3.5 H2O + 1.75 O2 = Fe(OH)3 + 2 SO4-2 + 4 H+
RATES
    Pyrite_oxidation
    -start
    10  rate = parm(1) * (m/m0)^0.67 * act("O2")^0.5
    20  save rate * time
    -end

One file = complete model: speciation + minerals + kinetics + transport. Close to ideal for our purposes.

Alternatives (same class, different flavour):

• GWB (Geochemists Workbench) — commercial, but similar text DSL.
• EQ3/6 (LLNL).
• CHNOSZ (R) — less DSL, more API.
• Cantera YAML with surface kinetics (≥3.0) — gas + electrochemical surface, close to our sulfide electrode context; lacks full aqueous speciation.

2.7. Databases and queries

• OPTIMADE 1.x — unified query DSL over Materials Project, AFLOW, OQMD, NOMAD, JARVIS, COD, Materials Cloud:

  elements HAS ALL "Fe","S" AND nelements=2 AND _mp_band_gap<0.1
  

• Materials Project schema — Pydantic models, readable as typed Python.
• NOMAD Metainfo — common schema for DFT results, parsers for ~40 codes.

2.8. Force fields / MLIP

• OpenKIM API — standard interface for classical potentials.
• MACE / CHGNet / SevenNet / UMA / MatterSim — not DSLs, but all share the extxyz format + OpenKIM-style metadata. For foundation MLIPs this is the de facto data-transfer language (energy/forces/stress in comments).
• SMIRNOFF — formally for organics, but Open Force Field is expanding to metal centres (2025).

2.9. Mineralogy

• IMA mineral list — names, not a language.
• End-member formula syntax (Berman, Helgeson notation) — for mineral thermodynamics.

2.10. Generative / LLM-friendly (2024–2026)

• CrystaLLM (Antunes et al., Nat Commun 2024) — GPT-style LM on CIF: provide composition + space group, receive a complete CIF. Perovskites/spinels/zeolites/MOFs.
• CrystalTextLLM (Meta) — Llama-2-70B fine-tuned on textual point-cloud descriptions.
• CrysText — combination of CIF + Llama 3.1-8B + QLoRA.
• Robocrystallographer (Materials Project) — inverse direction: CIF → readable English description (“face-sharing octahedra…”).
• Generative AI for crystals review (npj Comp Mat 2025) — review of the entire field.
• MatSci-LLM-Agents — agentic framework for materials science.

---

Sources

Part 1 (generic / organic)

• An Update to the SBML Human-Readable Antimony Language (arXiv 2405.15109)
• Antimony 3: Extending human-readable model definitions (bioRxiv 2026)
• MobsPy: A programming language for biochemical reaction networks (PLOS Comp Bio 2025)
• Antimony Reference (Tellurium docs)
• A High-Level Language for Rule-Based Modelling — Kappa (PLOS One)
• NOMAD Metainfo docs
• FAIRmat custom workflows tutorial
• LangSim — LLM Interface for Atomistic Simulation (MPIE)
• Foundation models for atomistic simulation of chemistry and materials (Nat Rev Chem 2025)
• 32 examples of LLM applications in materials & chemistry (IOP 2025)
• General-Purpose Models for the Chemical Sciences (Chem Reviews 2025)
• LLM-guided automated reaction pathway exploration (Comm Chem 2025)
• A multimodal LLM for materials science (Nat Mach Intell 2026)

Part 2 (inorganic / materials science)

• The CALPHAD method and its role in material development (PMC 4912057)
• CALPHAD + MLIP — Pt-W system case study (arXiv 2508.01028)
• Accelerating CALPHAD with universal MLIPs (arXiv 2411.15351)
• PHREEQC Version 3 (USGS)
• PHREEQC v3 input description (USGS TM 6-A43)
• Crystal structure generation with autoregressive LLM — CrystaLLM (Nat Commun 2024)
• CrystaLLM repository (lantunes)
• Crystal-text-LLM (Meta Research)
• Generative AI for crystal structures — review (npj Comp Mat 2025)
• LLMs Are Innate Crystal Structure Generators (Cornell 2025)
• A survey of AI-supported materials informatics (ScienceDirect 2025)
• MatSci-LLM-Agents (GitHub)

Additional (not from search, widely known)

• IUCr CIF dictionary — iucr.org/resources/cif
• AFLOW Prototype Encyclopedia — aflowlib.org/prototype-encyclopedia
• Bilbao Crystallographic Server / MAGNDATA — cryst.ehu.es
• pymatgen — pymatgen.org
• ASE — ase-lib.org (was wiki.fysik.dtu.dk/ase)
• AiiDA — aiida.net
• OpenKIM — openkim.org
• Open Force Field Initiative / SMIRNOFF — openforcefield.org

This site is open source. Improve this page.