Dataset Downloads

The complete chemical reaction network (CRN) explored in this work is available as a structured PostgreSQL dataset: compounds, potential-energy-surface minima, transition states, and elementary reactions with ML reaction paths and PBE0/def2-TZVPP refinement.

Reaction Atlas — full CRN
PostgreSQL 15 plain-SQL dump · gzip-compressed · ≈26 GB

Download from Zenodo →

The dataset is hosted on Zenodo (a persistent DOI is assigned on deposit).

Dataset at a glance

12,542

unique compounds

191

molecular formulas

33,659

PES minima

24,172

intra-PES TSs

47,987

elementary TSs

22,315

unique reactions

Compounds span C0–C4 backbones and charges from −3 to +2. For every reaction the dataset includes the full ML-predicted reaction path (TS, IRC trajectory, and connected reactant/product minima). For every reactive TS it includes PBE0/def2-TZVPP recomputed barriers, absolute energies of the TS and connected minima, and the Hessian. Where DFT indicates the ML-predicted TS is invalid, the corrected PBE0 TS, its energy difference, and the RMSD to the ML prediction are provided. PBE0 energies are additionally supplied with Grimme D3(BJ) and D4 dispersion corrections.

Restoring the database

The dump restores into a PostgreSQL 15 database. Create an empty database and stream the dump into it:

# 1. Create a database
createdb crn_atlas

# 2. Restore (gzipped plain-SQL dump)
gunzip -c reaction-atlas-*.sql.gz | psql -v ON_ERROR_STOP=1 -d crn_atlas

# The restored database is ≈26 GB including indexes — allow ample time and disk.

Decoding the binary columns

Geometry and Hessian columns are stored as PostgreSQL bytea holding serialized NumPy arrays. Single-array fields (positions, atomic numbers, Hessians) use numpy.save (.npy); trajectory fields use numpy.savez_compressed (.npz) with keys energies, positions (n_frames, n_atoms, 3), forces, and per-frame hessian_i.

import io, numpy as np

# Single array (e.g. minima.positions, reactions.ts_hessian_pbe0):
positions = np.load(io.BytesIO(row_bytes))          # -> (n_atoms, 3)

# Trajectory (e.g. reactions.reactant_trajectory):
npz = np.load(io.BytesIO(traj_bytes))
frames = npz["positions"]                            # -> (n_frames, n_atoms, 3)
energies = npz["energies"]

Schema reference

The public dataset contains the following tables. Operational tables used only during exploration (job queues, worker heartbeats, kinetics snapshots, UI annotations) are excluded. A few columns hold internal variants — alternative barrier references and manual-equilibrium overrides — that can be ignored for most analyses.

compounds

One row per unique chemical species (canonical SMILES). 12,542 rows.

Column	Type	Description
id	int PK	Primary key.
smiles	text	Canonical SMILES (unique).
formula	text	Molecular formula, e.g. C4H6O2.
charge	int	Net molecular charge (−3 … +2).
n_atoms	int	Number of atoms.
sorted_atomic_numbers	bytea (.npy)	Sorted atomic numbers of the compound.
is_seed	bool	Whether the compound was part of the initial seed set.
energy_pbe0	float	PBE0/def2-TZVPP single-point energy at the lowest-energy minimum (reference energy). Nullable.
energy_pbe0_method	text	Method/basis string, e.g. 'PBE0/def2-TZVPP'. Nullable.
energy_d3bj	float	Grimme D3(BJ) dispersion correction to the PBE0 energy (Hartree). Nullable.
energy_d4	float	Grimme D4 dispersion correction to the PBE0 energy (Hartree). Nullable.

minima

PES minima (conformers). 33,659 rows, averaging 2.68 conformers per compound.

Column	Type	Description
id	int PK	Primary key.
compound_id	int FK	→ compounds.id.
local_id	int	Sequential conformer index within the compound (0, 1, 2 …).
positions	bytea (.npy)	Cartesian coordinates, shape (n_atoms, 3).
energy	float	ML potential energy of the minimum.
hessian	bytea (.npy)	Mass-independent Hessian, shape (3n, 3n). Nullable.
explored	bool	Whether this minimum was expanded during exploration.
n_merged	int	How many raw minima were deduplicated into this one.
energy_pbe0	float	PBE0/def2-TZVPP single-point energy (populated for the lowest-E minimum). Nullable.
energy_d3bj	float	D3(BJ) dispersion correction to the PBE0 energy. Nullable.
energy_d4	float	D4 dispersion correction to the PBE0 energy. Nullable.

intra_transition_states

Intra-PES transition states connecting two conformers of the same compound. 24,172 rows.

Column	Type	Description
id	int PK	Primary key.
compound_id	int FK	→ compounds.id.
local_id	int	Sequential index within the compound.
positions	bytea (.npy)	TS Cartesian coordinates, shape (n_atoms, 3).
energy	float	ML potential energy at the TS.
eigenvalue	float	Imaginary-mode Hessian eigenvalue.
hessian	bytea (.npy)	Hessian at the TS, shape (3n, 3n). Nullable.
min_fwd_id / min_bwd_id	int FK	→ minima.id: the two conformers this TS connects.
barrier_fwd / barrier_bwd	float	ML barrier heights to the forward / backward minimum.
rmsd_to_fwd_min / rmsd_to_bwd_min	float	RMSD from the TS to each connected minimum.
fwd_trajectory / bwd_trajectory	bytea (.npz)	Relaxation paths from the TS to each minimum. Nullable.

reactions

Elementary chemical reactions (47,987 reactive TSs → 22,315 unique reactions by reactant–product identity). Includes the full ML reaction path, PBE0 refinement, and D3(BJ)/D4 dispersion corrections.

Column	Type	Description
id	int PK	Primary key.
ts_id	bigint	Global transition-state id (unique).
ts_conformer_positions	bytea (.npy)	TS geometry of the reacting supersystem, shape (n_atoms, 3).
ts_conformer_atomic_numbers	bytea (.npy)	Atomic numbers for the TS supersystem.
ts_conformer_charge	int	Total charge of the reacting supersystem.
ts_energy	float	ML potential energy at the TS.
barrier_forward / barrier_backward	float	In-box ML barriers (TS − trajectory endpoint).
barrier_forward_separated / …_backward_separated	float	Separated ML barriers (TS − sum of reference-conformer energies). Nullable.
reactant_trajectory / product_trajectory	bytea (.npz)	ML IRC trajectories to the reactant / product side. Nullable.
energy_R_pbe0 / energy_TS_pbe0 / energy_P_pbe0	float	PBE0/def2-TZVPP absolute energies of reactant, TS, product. Nullable.
barrier_forward_pbe0 / barrier_backward_pbe0	float	In-box PBE0 barriers. Nullable.
barrier_forward_separated_pbe0 / …_backward…	float	Separated PBE0 barriers (primary for kinetics). Nullable.
energy_pbe0_method	text	PBE0 method/basis string. Nullable.
ts_hessian_pbe0	bytea (.npy)	PBE0 Hessian on the TS geometry, shape (3n, 3n). Nullable.
ts_hessian_ml	bytea (.npy)	ML Hessian on the TS geometry, shape (3n, 3n). Nullable.
ts_ml_invalid	bool	True when DFT shows the ML TS is not a true saddle (≠ exactly one imaginary mode).
ts_pbe0_corrected_positions	bytea (.npy)	Corrected DFT TS geometry (PBE0 saddle-point opt). Populated only when ts_ml_invalid. Nullable.
ts_pbe0_corrected_energy	float	PBE0 energy of the corrected TS. Nullable.
ts_pbe0_corrected_de	float	E(corrected) − E(ML at ML geometry). Nullable.
ts_pbe0_corrected_rmsd	float	Kabsch-aligned RMSD of corrected TS vs ML TS. Nullable.
energy_R_d3bj / energy_TS_d3bj / energy_P_d3bj	float	D3(BJ) dispersion corrections to the reactant / TS / product PBE0 energies. Nullable.
energy_R_d4 / energy_TS_d4 / energy_P_d4	float	D4 dispersion corrections to the reactant / TS / product PBE0 energies. Nullable.
ts_corrected_d3bj / ts_corrected_d4	float	Dispersion corrections on the corrected DFT TS. Nullable.
barrier_*_pbe0_d3bj	float	PBE0+D3(BJ) dispersion-corrected barriers (forward / backward, in-box and separated). Nullable.
barrier_*_pbe0_d4	float	PBE0+D4 dispersion-corrected barriers (forward / backward, in-box and separated). Nullable.
disp_method	text	Dispersion-correction method label. Nullable.

reaction_reactants

Reactant side of each reaction (compound + conformer).

Column	Type	Description
id	int PK	Primary key.
reaction_id	int FK	→ reactions.id.
compound_id	int FK	→ compounds.id (a reactant species).
conformer_local_id	int	Which conformer (minima.local_id) participates. Nullable.

reaction_products

Product side of each reaction (compound + conformer).

Column	Type	Description
id	int PK	Primary key.
reaction_id	int FK	→ reactions.id.
compound_id	int FK	→ compounds.id (a product species).
conformer_local_id	int	Which conformer participates.
energy	float	Energy of the product conformer.

graph_edges

Materialized bipartite edges of the reaction network (compound ↔ reaction nodes) for the graph view. 129,969 rows.

Column	Type	Description
id	int PK	Primary key.
source_node / target_node	text	Node identifiers (compound or reaction).
source_type / target_type	text	'compound' or 'reaction'.
direction	text	Edge direction / role. Nullable.
stoichiometry	int	Stoichiometric coefficient.
energy_diff	float	Energy difference along the edge. Nullable.
reaction_id	int FK	→ reactions.id. Nullable.

License & citation

Please cite the accompanying paper when using this dataset; full citation details and the license are provided on the Zenodo record.