Frequently Asked Questions

Note

Explanation (why the format looks like this). Learning path: Tutorial — your first CON checkpoint. Task recipes: How-to — CON I/O by language.

What CON is for

CON stores one atomic configuration that rare-event codes already depend on. A frame keeps the cell, type-grouped coordinates, a per-direction constraint mask, a pre-group atom_id, optional per-atom sections, and line-2 JSON. Tools that invent a private dump lose the mask or the identity on the first rewrite.

On disk	Role
Cell + angles	Periodic box
Type-grouped coordinates	Stable head-able layout
Column 4 fixed mask	Per-direction constraints (bitmask 0-7)
Column 5 atom_id	Pre-group index for NEB / dimer / reference matching
Optional sections	Velocities, forces, energies, charges, spins, magmoms, displacements, spreads (v2/v3 sections)
Line-2 JSON	con_spec_version, energy, neb_bead, units, …

Saddle, dimer, and NEB pipelines already depend on that payload. readcon-core is the spec v2-v3 reader/writer and the hourglass rkr_* ABI in each language. Chemfiles converts into CON; DLPack and metatensor export out of it. index_proj and readcon-db index corpora that stay CON text. Spec: The CON File Format Specification. Design history: Format Evolution and Design Rationale.

Is frame topology (bonds) required?

No. bonds is an optional v2 metadata key (not a sections block and not a CON spec v3 change). Legacy files omit it. When present, each entry is a 0-based atom_data index pair (optionally with chemfiles-style order). It enables tools such as chemfiles selection (bonds: / angles: / is_bonded) when the library is built with --features chemfiles.

How do I select atoms (and optional bonds) on a CON frame?

Selection is a CON API: strings such as name H or bonds: all return atom_data indices (or structured matches) via select_on_frame / select_atom_indices / rkr_frame_select and language wrappers. name / type / all need only symbols. Topology selectors need metadata["bonds"] (hand-authored or filled when converting into CON). All languages share one evaluator.

The con format has no residue table, stores pair bonds only, and keeps a thin optional property subset after foreign-format import. So resname, most external property maps, impropers, and geometry-threshold minidialects are not on the selection surface: there is simply no data for them. Detail: chemfiles-explain.org (What selection cannot see on CON).

Foreign formats (XYZ, PDB, …) enter through an optional conversion feature (Cargo chemfiles, PyPI readcon-chemfiles). Lean builds still expose the selection symbols but error clearly until that feature is linked. Docs for conversion live under docs/orgmode/chemfiles-*.org; language APIs in bindings.org; on-disk bonds in spec.org.

On import only, display names such as H1 versus types such as H may be kept in sidecars (chemfiles_atom_names / chemfiles_atom_types); the on-disk column remains a single symbol. Hand-built frames without sidecars use symbol for both.

What problems does atom_id solve?

The con format groups atoms by element type. A structure with atoms C, C, C, O, C, C (indices 0-5) gets written as five C atoms followed by one O atom. Without a persistent identity field, the original ordering vanishes after one read-write cycle.

This matters for:

• NEB calculations: interpolated images must maintain consistent atom ordering across the band. If atom ordering drifts, the interpolation produces nonsense.

• Dimer searches: the displacement vector references specific atom indices. Reordering atoms invalidates the mode.

• Reference comparisons: comparing a relaxed structure against a reference (e.g., Baker test set) requires matching atoms by index.

The atom_id field (column 5) stores the pre-grouping index, allowing exact reconstruction of the original ordering after any number of read-write cycles.

Why JSON on line 2?

Line 2 was historically unused (“Time” or empty in eOn files). JSON provides:

• Forward compatibility: new keys can be added without format changes. Unknown keys are preserved through round-trips.

• Machine readability: no custom parser needed. Every language has a JSON library.

• Section declaration: the sections key tells the parser exactly what per-atom data to expect, eliminating ambiguity.

• Provenance: potential, energy, generator keys make files self-documenting.

• Backward compatibility: pre-v2 files have non-JSON on line 2. The parser detects this (line 2 does not start with {) and falls back to legacy mode (spec_version = 1).

How fast is readcon-core?

CON parse uses fast-float2, zero-copy line views, header-sized vectors, optional mmap, and forward / forward_fast when you only need counts. CI: Cachegrind I-refs (ci_cachegrind.yml). PRs: Python ASV + spyglass (ci_benchmark.yml, suite in benchmarks/). Local peer scripts: benches/compare_readers.py, multiformat_traj.py, ase_traj_vs_con.py, h5md_vs_con.py. How to re-run and what each gate means: Performance Benchmarks.

What is the sections mechanism?

Version 2+ files can include per-atom data beyond coordinates. Each additional section follows the same block structure as coordinates: blank separator, symbol line, label line, data lines.

The sections key in the JSON metadata declares which sections exist and their order. Known names on the v2/v3 surface:

Name	Layout
velocities, forces, magmoms, displacements, spreads	3-vector + fixed + atom_id
energies, charges, spins	scalar + fixed + atom_id

{"con_spec_version":2,"sections":["velocities","forces","charges"]}

Optional physics blocks such as charges / spins / magmoms / displacements / spreads use the same declared-section channel; they do not require a new con_spec_version.

Compared with the legacy approach (detecting velocities by peeking for a blank separator):

• Declared sections fix what the reader must consume

• New section names need no positional guesswork once reserved in the spec

• Section order is explicit

Declared sections must be present at the declared position. Use an empty sections array ([]) to state that no additional per-atom sections follow. When the key is absent, the reader keeps the blank-separator fallback for existing .convel files.

Legacy .convel files without a sections key still work: the parser falls back to blank-separator velocity detection.

What does validate=true do?

The validate metadata key asks v2 readers to reject frames that do not satisfy strict ordering and schema invariants:

{"con_spec_version":2,"sections":["velocities"],"validate":true}

In this mode, sections must be present. Readers verify the declared section order, exact component labels, component symbols, integer identity columns, matching fixed masks and atom ids across sections, finite numeric values, physical cell geometry, positive counts and masses, and the JSON types of reserved metadata keys.

Can I store forces, energies, charges, spins, magmoms, displacements, spreads?

Yes. Per-frame total energy lives in JSON under the energy key. Per-atom data uses declared sections:

• Forces (3-vector): sections includes forces

• Per-atom energies (scalar): energies

• Charges / spins (scalar): charges, spins

• Magnetic moments (3-vector): magmoms

• Displacement vectors (3-vector, Angstrom, e.g. a normal mode): displacements

• Spreads (3-vector, Angstrom): spreads, the root-mean-square spread of each atom along x, y, z about its written coordinates; a standard deviation, 0 for a classical point

{"con_spec_version":2,"sections":["forces"],"energy":-42.5,"potential":{"type":"EMT","params":{"cutoff":6.0}}}

For ML potentials that decompose total energy into per-atom contributions, declare energies alongside forces:

{"con_spec_version":2,"sections":["forces","energies"],"energy":-42.5}

Charges, spins, magmoms, displacements, and spreads use the same wire format: list them in sections and emit the matching component blocks (see The CON File Format Specification). Example fixtures: resources/test/tiny_cuh2_charges_spins_magmoms.con, resources/test/tiny_cuh2_displacements.con, resources/test/tiny_cuh2_spreads.con.

The per-frame energy metadata key SHOULD equal the sum of the per-atom energies section when both are present. Frames with forces SHOULD identify the potential (potential key) so downstream tools know how to interpret the values.

Does readcon-core support compression?

Yes. Two formats are detected automatically by magic bytes:

• gzip: .con.gz extension or 0x1f 0x8b magic; always available.

• zstd: .con.zst extension or 0x28 0xb5 0x2f 0xfd magic; opt-in behind the zstd Cargo feature. Builds without the feature still detect zstd input and return a clear error pointing at the feature flag rather than producing a corrupt parse.

Writing through ConFrameWriter::from_path_gzip / from_path_gzip_with_precision or (with the zstd feature) the matching from_path_zstd constructors compresses output transparently.

Force data adds three floats per atom (roughly 3× the coordinate payload alone). Gzip and (with the zstd feature) zstd reduce on-disk size; measure on your trajectories rather than treating a single ratio as fixed.

How do I look up an atom by its atom_id?

readcon-core preserves atom_id (column 5) through every read-write cycle, but the in-memory atom order follows the file’s type-grouped layout, not the atom_id ordering. Two convenience APIs lift the gap:

• One-shot: frame.atom_index_by_id(id) scans the atom list and returns Option<usize>. O(N) per call.

• Repeated: frame.build_atom_id_index() returns an FxHashMap<u64, usize> (Rust) / dict (Python) / Dict{UInt64, Int} (Julia) for O(1) reverse lookup. Build once and reuse for every lookup against the same frame.

Both APIs mirror across every supported binding (Rust, C ABI, C++, Python, Julia).

What languages are supported?

Language	Mechanism	Installation
Rust	Native crate	cargo add readcon-core
Python	PyO3 bindings	pip install readcon
C	FFI (cdylib)	link libreadcon_core
C++	RAII header	#include "readcon-core.hpp"
Julia	ccall wrapper	using ReadCon

All bindings share the same Rust core, ensuring identical parsing behavior across languages.

How do I convert between ASE and con?

import readcon

# con -> ASE Atoms (preserves atom_id, velocities, forces, masses)
frames = readcon.read_con("input.con")
ase_atoms = frames[0].to_ase()

# ASE Atoms -> con
frame = readcon.ConFrame.from_ase(ase_atoms)
readcon.write_con("output.con", [frame])

# Direct read to ASE list
ase_list = readcon.read_con_as_ase("trajectory.con")

The conversion preserves atom_id (via a custom per-atom array), velocities, forces (via SinglePointCalculator), masses, and constraints (FixAtoms).

Why both readcon-core.h and metatensor.h / readcon-metatensor.h?

readcon’s cbindgen surface and metatensor-sys’s cbindgen surface are separate crates. We hand off opaque mts_block_t *; values and labels use metatensor’s C API. Prefer include/readcon-metatensor.h (metatensor.h first). Lean builds still export metatensor entry points that return RKR_STATUS_FEATURE_DISABLED (-11), distinct from internal error (-7).

Lean vs fat libreadcon_core for C/Fortran?

Same symbol names either way. Without --features metatensor, blocks return -11. Without --features zstd, create_writer_zstd_* returns a null writer. gzip writers and DLPack (ArcArray share via dlpk; no fake _borrowed C aliases) are always present. After a metatensor-enabled build, target/<profile>/readcon-metatensor.env lists include/lib paths for libmetatensor.

What is the stack for?

One on-disk checkpoint and one rkr_* ABI, instead of each code hand-rolling XYZ and its own atoms type. Migration guide: How-to — migrate a stack onto CON (readcon-core convert, readcon.convert_to_con).

Component	Job
CON on disk	The checkpoint format (text, optional gzip/zstd)
readcon-core	Frame API + hourglass ABI + chemfiles in + selection + compression + DLPack/metatensor (docs.rs)
readcon-db	LMDB corpus: energy / formula / section indexes, dedup, multi-reader (docs)
Chemfiles	Land foreign structures as CON
ASE adapters	Calculators without abandoning CON interchange
chemparseplot / rgpycrumbs	Plotting and analysis on CON checkpoints
rgpot / eOn	Optimizers and potentials on the same CON files

One on-disk format. One library API. Campaigns, selection, and plotting share it. Campaign field projection helpers live in `index_proj <https://docs.rs/readcon-core/latest/readcon_core/index_proj/>`_ (same meanings readcon-db indexes).

Are ASE adapters the primary API?

No. Optional to_ase / from_ase support calculator hand-off. Interchange and multi-reader campaign storage use readcon / readcon-db with CON text authoritative.

What is authoritative: CON text, cooked SoA, or DLPack?

UTF-8 CON text is authoritative for files and for readcon-db frames (hash, dedup, join/split, reindex). Cooked RCSO / SoA / DLPack buffers accelerate extract and device hand-off and are discardable or ephemeral.

Why is readcon-db a separate package?

readcon-core is the shared decoder/writer. readcon-db owns LMDB indexes and SWMR corpus access. That split is a migration benefit: once structures are CON text, the same files plug into corpus query (energy / formula / section indexes, dedup) without rewriting the optimizer or potential. Install readcon-db separately (cargo add readcon-db, pip install readcon-db). Package docs: lode-org.github.io/readcon-db/docs. Source: github.com/lode-org/readcon-db.

Where do large campaigns and many frames go?

In this stack: CON text remains the structure contract; readcon-db is the LMDB corpus on top of it (indexes for energy / formula / section presence, content-hash dedup, multi-reader SWMR). Multi-frame CON files (optionally gzip/zstd) and iter_con / forward cover trajectory-style loads in readcon-core itself. Large corpora stay CON text; readcon-db indexes them. Do not switch the on-disk structure to another dialect for screening.

Install: cargo add readcon-db / pip install readcon-db. Package docs: lode-org.github.io/readcon-db/docs.

How do XYZ, PDB, GRO, and chemfiles fit in?

They are ingress. Chemfiles (Cargo chemfiles, PyPI readcon-chemfiles) maps foreign structures into ConFrame / CON so the rest of the stack speaks one format. The durable interchange is CON (constraints, atom_id, sections, JSON, hourglass ABI). Convert at the edge when inputs arrive as XYZ/PDB/GRO; keep CON for optimizers, campaigns (readcon-db), selection, and plotting. How-to: How-to — migrate a stack onto CON, Convert other formats into CON.

Why an hourglass C ABI?

Optimizers and drivers are often Fortran or C++. A single rkr_* surface gives those codes the same CON semantics as Python and Julia without embedding a Python interpreter on the I/O path. The 0.14.x stability contract (ownership, sentinels, no-abort errors, Z=1..92 helpers) is on the architecture page. A 1.0.0 cut is the production classifier, not a silent ABI rewrite.