TRIP50

TRIP50 is a 50-reaction benchmark for radical and triplet thermochemistry and forward/reverse barriers, with a DLPNO-CCSD(T) reference and 46 DFT method totals computed on the same geometries. The benchmark is small, well-curated, and chemically diverse — designed to evaluate methods on relative energetics: reaction energies (ΔE_rxn), forward barriers (ΔE‡_fwd), and reverse barriers (ΔE‡_rev).

This release packages the structures and reference energies as a single Hugging Face dataset for use evaluating machine-learning interatomic potentials (MLIPs) alongside conventional DFT.

No forces. TRIP50 is a single-point energy benchmark on relaxed stationary points. MLIPs that require force training data should look elsewhere; MLIPs being evaluated on relative energetics work directly.

Contents

File	Rows	Notes
data/trip50.extxyz	156 frames	ASE-readable, all 47 method energies on the comment line
data/species.parquet	156	one row per canonical structure, full float64 precision
data/reactions.parquet	50	reaction definitions with reference Δs in kcal/mol
data/aliases.parquet / aliases.csv	30	logical alias map (no duplicate xyz files shipped)
data/methods.json	–	slug ↔ display-name mapping for the 47 method columns
data/MANIFEST.sha256	–	digest of every payload file

Schemas

species.parquet

column	type	notes
species_id	string	canonical name, e.g. 1-R1, 26-TS
rxn_id	int32	1..50
role	string	R1 / R2 / TS / P1 / P2
n_atoms	int32
atomic_numbers	list<int8>	length n_atoms
positions	list<list<float64, 3>>	Cartesian, Å
charge	int8	always 0
multiplicity	string	singlet / doublet / triplet
spin_multiplicity	int8	1 / 2 / 3
energy_dlpno_ccsd_t	float64	reference, Hartree
energy_<slug> × 46	float64	DFT methods, Hartree (slugs in methods.json)

reactions.parquet

column	type	notes
rxn_id	int32	1..50
category	string	one of C-C, C-O, C-S, HAT, Si-X, C-Hal, N-X (Figure 2 of the source paper)
r1_species_id, r2_species_id, ts_species_id, p1_species_id, p2_species_id	string (nullable)	post-alias canonical IDs; null = unimolecular leg (contributes 0 H)
is_unimolecular_reactant, is_unimolecular_product	bool
dE_rxn_kcal_dlpno	float64	E(P1)+E(P2) − E(R1)−E(R2), kcal/mol
dE_fwd_kcal_dlpno	float64	E(TS) − E(R1)−E(R2), kcal/mol
dE_rev_kcal_dlpno	float64	E(TS) − E(P1)−E(P2), kcal/mol

aliases.parquet

column	type	notes
alias	string	non-canonical species id seen in the rxn definitions
canonical	string	the species_id actually present in species.parquet

Aliases are logical only — the dataset ships one xyz per canonical structure. A consumer evaluating reaction 50 should look up p1_species_id = "48-R1" (which is the same molecule as 50-P1).

Conventions

• Units. Total energies in species.parquet and trip50.extxyz are in Hartree. Reference reaction-level Δ values in reactions.parquet are in kcal/mol (1 Hartree = 627.5094740631 kcal/mol).
• Charge & spin. Every species is neutral. Multiplicities span singlet / doublet / triplet. Spin-unaware MLIPs will incur unavoidable error on the radical and triplet species — a known limitation of the dataset for such models.
• Unimolecular legs. Reactions with a single reactant or product have the corresponding R2 / P2 field set to null in reactions.parquet, and the missing partner contributes 0 H to ΔE sums. This convention is inherited from the source paper.
• Alias resolution. Always go through reactions.parquet (which already stores post-alias canonical IDs). Never assume {rxn}-{role}.xyz exists for every (rxn, role).
• Precision. The Parquet artefacts are full float64. The trip50.extxyz representation rounds positions to ASE's default 8-decimal format (≈ 5 × 10⁻⁹ Å — far below physical precision). Use species.parquet if exact byte-identity matters.

Usage

Hugging Face datasets

from datasets import load_dataset

species = load_dataset("patonlab/trip50", "species", split="train")
reactions = load_dataset("patonlab/trip50", "reactions", split="train")

print(species[0]["species_id"], species[0]["energy_dlpno_ccsd_t"], "Hartree")
print(reactions[0]["dE_fwd_kcal_dlpno"], "kcal/mol")

ASE

from ase.io import read

frames = read("hf://datasets/patonlab/trip50/data/trip50.extxyz", index=":")
e_dlpno = frames[0].info["energy_dlpno_ccsd_t"]   # Hartree
mult = frames[0].info["multiplicity"]              # 'singlet' | 'doublet' | 'triplet'

A reference MLIP evaluator that prints a per-method MAE table is in examples/evaluate_mlip.py.

Limitations

• No forces or gradients. Relative reaction energetics only.
• Single geometries. No conformer ensembles; no thermochemical corrections (ΔH / ΔG); no solvent corrections.
• Reference is DLPNO-CCSD(T), not experiment. Methods that systematically agree with the reference may still disagree with experiment.
• Spin-aware models recommended. Roughly half the species are open-shell.

Citation

Please cite the original TRIP50 publication when using this dataset. See CITATION.cff for a machine-readable record.

@article{trip50,
  title   = {Fundamental Study of Density Functional Theory Applied to Triplet State
             Reactivity: Introduction of the {TRIP50} Data Set},
  author  = {Hughes, William B. and Popescu, Mihai V. and Paton, Robert S.},
  journal = {Journal of Chemical Theory and Computation},
  volume  = {22},
  pages   = {3530--3542},
  year    = {2026},
  doi     = {10.1021/acs.jctc.6c00144}
}

License

Creative Commons Attribution 4.0 International (CC-BY-4.0).