QuantumChem-200K: Photoinitiator Screening Dataset

• QuantumChem-200K is a large-scale dataset of over 200K organic molecules annotated with 11 key quantum-chemical and screening properties tailored for TPP photoinitiator discovery.
• It consolidates QM9-derived and OMG-derived molecules to merge ground-state and excited-state descriptors, including TPA and ISC, for enhanced mechanistic insights.
• The hybrid workflow combines DFT, semi-empirical methods, neural networks, and cheminformatics tools to enable accurate AI-assisted property prediction.

Searching arXiv for the named dataset and closely related comparison papers. arXiv search query: "QuantumChem-200K" QuantumChem-200K is a large-scale open organic molecular dataset curated for quantum-chemistry property screening in photoinitiator discovery, especially for two-photon polymerization (TPP). It contains more than 200,000 organic molecules annotated with eleven quantum-chemical and practical screening properties, including two-photon absorption (TPA) quantities, singlet–triplet intersystem crossing (ISC) energy, toxicity, synthetic accessibility, hydrophilicity, solubility, boiling point, molecular weight, and aromaticity. Its stated purpose is to supply the property coverage needed for data-driven screening and AI-assisted design of photoinitiators, and to support language-model benchmarking for forward property prediction from SMILES (Zeng et al., 23 Nov 2025).

1. Dataset identity and scientific setting

QuantumChem-200K was introduced to address a specific gap in open molecular resources: standard datasets such as QM7, QM9, GDB, and ZINC20 generally provide only basic molecular descriptors or ground-state properties, whereas photoinitiator screening for TPP requires photophysical and excited-state information. The dataset is therefore organized around properties that are mechanistically relevant to photodissociation-driven initiation, rather than around equilibrium electronic-structure observables alone (Zeng et al., 23 Nov 2025).

The motivating application is TPP under near-infrared excitation. The paper describes the sequence as absorption of one or two photons, promotion from the ground state $S_0$ to the first excited singlet state $S_1$, intersystem crossing to the triplet state $T_1$, and then bond cleavage for Norrish Type I photoinitiators, producing radicals that initiate polymerization. Within that framing, a candidate photoinitiator should ideally combine a large TPA cross section, especially near $780\,\mathrm{nm}$, with a small $S_1$-$T_1$ gap and acceptable formulation and developability properties.

A recurrent point in the paper is that photodissociation quantum yield would be the most direct screening target, but is difficult to compute reliably and labor-intensive to measure experimentally. QuantumChem-200K is consequently constructed as a surrogate screening resource: it aggregates proxy properties that are intended to be informative for photoinitiator performance without claiming to measure photodissociation yield directly.

2. Corpus construction and chemical scope

The dataset is assembled from approximately 210K source SMILES and results in a final corpus described as over 200,000 molecules. It merges two open molecular sources with different chemical roles: a QM9-derived small-molecule subset and an Open Macromolecular Genome (OMG) monomer subset.

Source subset	Scale and composition	Property coverage
QM9-derived subset	about 134,000 molecules; up to 9 heavy atoms; CHNOF	all properties
OMG monomer subset	about 77,000 monomers; up to 25 heavy atoms; C, H, N, O, F, Br, Cl, Si, P, S	all except ISC

This split is scientifically consequential. The QM9-derived subset contributes optimized ground-state geometries and is especially important for the ISC workflow, because the semi-empirical excited-state computation used for ISC is limited to CHNO or CHNOF molecules. The OMG-derived subset broadens the chemical domain toward larger monomer-like organics with chromophores, electron-donating and electron-withdrawing motifs, bond-cleavable groups, and structures resembling known Norrish Type I photoinitiators (Zeng et al., 23 Nov 2025).

The effective filtering criterion emphasized in the paper is elemental compatibility with the excited-state solver. The authors also applied the same compatibility logic when constructing the external benchmark, retaining only CHNOF molecules from VQM24 for ISC-related evaluation. This means that the dataset is not chemically uniform in label availability: ISC is not universally available across the full collection.

The primary molecular representation is SMILES. At the same time, the workflow uses optimized molecular geometry for property computation where required. The dataset is publicly available at https://huggingface.co/YinqiZeng704, but the manuscript does not specify a canonical file format or a standard train/validation/test split for the released corpus itself.

3. Property system and mechanistic interpretation

The paper describes QuantumChem-200K as containing eleven key properties, but its practical organization is easiest to understand in terms of property families.

The first family is TPA. The dataset records the maximum TPA cross section across the computed spectrum, the TPA cross section at $780\,\mathrm{nm}$, and the wavelength range over which absorption is appreciable. The absorption window is defined operationally as the range where the TPA cross section exceeds $20$ GM. These quantities are central because TPP relies on two-photon excitation, and $780\,\mathrm{nm}$ is singled out as a standard near-infrared wavelength in that setting (Zeng et al., 23 Nov 2025).

The second family is ISC. The relevant descriptor is the singlet–triplet gap

$\Delta E_{\mathrm{ISC}} = E_{S_1} - E_{T_1}.$

The paper uses smaller $S_1$0 values as a mechanistic surrogate for more favorable intersystem crossing, and therefore as an indirect indicator of photodissociation promise in Norrish Type I systems.

The third family comprises practical screening descriptors. These include toxicity, synthetic accessibility, hydrophilicity expressed as logP, solubility, boiling point, molecular weight, and aromaticity. Toxicity and synthetic accessibility are mapped to values from $S_1$1 to $S_1$2; the text states that for toxicity, $S_1$3 denotes low toxicity and $S_1$4 high toxicity, and for synthetic accessibility, $S_1$5 denotes easy synthesis and $S_1$6 challenging synthesis. The paper notes an inconsistency between this interpretation and the summary-table arrow for synthetic accessibility, so the score is best treated as a difficulty-like quantity in the explanatory text.

Two clarifications are important. First, QuantumChem-200K is not described as a direct excited-state dynamics dataset, nor as a dataset of experimental photoinitiator performance. Second, the paper’s appendix discusses photodissociation quantum yield and bond dissociation energy for scientific motivation, but explicitly does not treat them as part of the eleven main annotations. A common misconception is therefore to read the dataset as a direct label set for photodissociation efficiency; the paper instead positions it as a proxy-property resource.

4. Computational workflow and annotation pipeline

The annotation workflow is explicitly hybrid, combining DFT, semi-empirical excited-state methods, an atomistic quantum platform, neural-network predictors, and cheminformatics tools. This methodological heterogeneity is central to the dataset’s scale and also to its interpretive limits (Zeng et al., 23 Nov 2025).

For ISC, the workflow is two-stage. Ground-state geometry optimization is performed using B3LYP with the 6-31G(2df,p) basis set. The first singlet and triplet excitation energies are then computed using MNDO(ODM2*) on the Atomicistic/AIQM1 platform. The ISC descriptor is computed from the resulting $S_1$7 and $S_1$8 values.

For TPA, the paper states that two-photon absorption cross sections are computed using the Atomicistic Quantum Platform, cited as MLatom3 / Atomicistic platform. The response is evaluated over

$S_1$9

with Et55.4 (1-octanol) as solvent. The stored outputs include maximum TPA cross section, TPA absorption window, and TPA cross section at $T_1$0.

Toxicity and synthetic accessibility are predicted with eToxPred, described as a neural network trained on experimentally annotated toxicity datasets and retrosynthetic complexity metrics. Additional descriptors are generated with JRgui, OpenBabel, and RDKit. The paper assigns boiling point, solubility, and hydrophilicity to JRgui; molecular weight to OpenBabel; and aromaticity to RDKit.

This workflow implies that the dataset is not a single-level ab initio corpus. A plausible implication is that different properties inherit different uncertainty structures and calibration regimes. The paper does not provide per-property confidence intervals, a detailed uncertainty model, or a large experimental calibration study, so the annotations are best interpreted as high-throughput computed or estimated labels rather than uniformly validated ground truth.

5. Language-model benchmarking and external evaluation

A second major function of QuantumChem-200K is as a benchmark and fine-tuning corpus for a chemistry LLM. The authors fine-tuned Qwen2.5-32B to create a chemistry AI assistant for forward property prediction from SMILES, and evaluated it with a custom weighted mean absolute error, $T_1$1, which normalizes for property range and reweights by property availability (Zeng et al., 23 Nov 2025).

The paper does not define a canonical split for the dataset release, but it does define an external benchmark consisting of 3000 unseen molecules. This testbank includes 1000 molecules from VQM24, restricted to CHNOF for ISC compatibility, and 2000 molecules from ZINC20, described as drug-like structures with up to 20 heavy atoms and element diversity including CHNOFBrSiSCl. The authors state that the same automated workflow used for QuantumChem-200K was applied to compute all eleven properties for these molecules.

On internal monitoring over 100 randomly sampled QuantumChem-200K molecules, the paper reports that overall wMAE decreased by about 30% between 3 epochs and 6 epochs of fine-tuning. For TPA at 780 nm, wMAE improved from 0.027 to 0.017; for ISC, from 0.023 to 0.017.

On the 3000-molecule unseen testbank, the fine-tuned assistant achieved overall wMAE = 0.1975. The reported baselines are:

• Fine-tuned AI assistant: 0.1975
• Llama-3.1-70B: 2.2195
• Qwen2.5-32B: 3.3038
• GPT-4o: 3.9200

The paper highlights performance on photoinitiator-relevant quantities in particular. On the external testbank, TPA $T_1$2 at $T_1$3 reached wMAE 0.011, and maximum TPA $T_1$4 reached wMAE 0.012. The authors also state that ISC prediction accuracy remained strong and exceeded the three unfine-tuned baseline LLMs, although the prose does not quote the exact external ISC wMAE value.

6. Relation to prior resources, limitations, and interpretation

QuantumChem-200K is positioned against both standard molecular datasets and broader quantum-chemistry resources. Relative to QM7, QM9, GDB, and ZINC20, its distinctiveness lies in annotating photoinitiator-relevant excited-state and practical screening properties rather than only basic descriptors. Relative to resources such as ANI-1ccx, VQM24, and PubChemQC, the paper argues that the dataset supplies a specific combination of TPA, ISC, and deployability-related annotations that is tailored to TPP screening (Zeng et al., 23 Nov 2025).

It is also distinct from conformer-rich DFT corpora aimed at potential-energy-surface learning. QO2Mol, for example, is an open dataset of 120,000 ChEMBL-derived organic fragments with approximately 20 million conformers, energies, and forces at $T_1$5, and is designed for neural network potentials and machine-learning force fields rather than photoinitiator screening or SMILES-based LLM benchmarking (Liu et al., 2024). This contrast helps situate QuantumChem-200K within the larger landscape: it is property-centric and application-specific, not a general-purpose conformer database.

Several limitations follow directly from the paper. The most important is the surrogate-property limitation: the dataset does not annotate photodissociation quantum yield, which the authors themselves identify as the most informative photoinitiator-performance indicator. Instead, it supplies TPA, ISC gap, and practical descriptors as proxies. Label availability is uneven because ISC depends on MNDO(ODM2*) compatibility. The hybrid pipeline introduces cross-method bias, and the manuscript does not provide explicit uncertainty estimates, calibration curves, or a detailed error decomposition for the underlying property engines.

A further limitation is representational. The chemistry assistant predicts from SMILES alone, whereas several target properties, especially toxicity and excited-state behavior, may depend strongly on geometry, environment, and electronic structure. The paper explicitly notes that toxicity remains particularly difficult to predict and is the largest error contributor. This suggests that QuantumChem-200K should be read not as a final statement on photoinitiator efficacy, but as a large open screening substrate for multi-objective prioritization, surrogate modeling, and chemistry-aware LLM evaluation.