The QM9 dataset is a collection of roughly 130,000 small organic molecules whose quantum-mechanical properties have been computed and catalogued, making it one of the most widely used benchmarks in the field of machine learning for chemistry. Each molecule in QM9 contains up to nine “heavy” atoms (carbon, nitrogen, oxygen, and fluorine, plus hydrogen) and comes with a dozen calculated ground-state properties spanning geometry, energy, electronic structure, and thermodynamics. Its influence on the development of molecular machine learning is hard to overstate, but it also has real limitations that shape what researchers can and cannot learn from it.
What QM9 Actually Contains
QM9 encompasses 133,885 molecules whose properties were computed using density functional theory, a workhorse method in computational chemistry that approximates the behavior of electrons in a molecule. The specific flavor used was B3LYP with a particular basis set, giving each molecule a consistent set of twelve ground-state properties: things like total energy, the energies of the highest occupied and lowest unoccupied molecular orbitals (HOMO and LUMO), the gap between those orbitals, dipole moment, polarizability, zero-point vibrational energy, heat capacity, and more.1Journal of Cheminformatics. Dataset’s chemical diversity limits the generalizability of machine learning predictions The molecules themselves are drawn from a systematic enumeration of possible small organic structures, which means QM9 is not a random sample of chemistry but a near-exhaustive catalog of what you can build with a handful of light atoms.
The restriction to five elements and nine heavy atoms keeps the molecules small. You will find simple amines, alcohols, ethers, small ring systems, and short chains, but nothing remotely resembling a drug molecule or a polymer. That deliberate simplicity is part of why the dataset became so popular: it is clean, consistent, and manageable. Researchers describe it as homogeneous, pure, and largely free of the noise that plagues real-world chemical data.1Journal of Cheminformatics. Dataset’s chemical diversity limits the generalizability of machine learning predictions
Why Electronic Properties Matter So Much
Of all the properties in QM9, the HOMO energy, LUMO energy, and the gap between them get the most attention from researchers. These values describe how tightly a molecule holds onto its electrons and how readily it can accept new ones. In practical terms, the HOMO-LUMO gap influences how a molecule absorbs light, conducts electricity, and reacts with other molecules. This makes it central to designing solar cells, organic LEDs, semiconductors, and pharmaceuticals.
Machine learning models trained on QM9 have gotten remarkably good at predicting these electronic properties from molecular structure alone. One recent stacking ensemble model achieved near-perfect accuracy on HOMO and LUMO predictions, with root-mean-square errors on the order of a few ten-thousandths of a Hartree (the standard energy unit in quantum chemistry).2PubMed Central. Predicting electronic properties of molecules: a stacking ensemble model for HOMO and LUMO energy estimation Other approaches have shown that even simple text-based molecular representations, without any three-dimensional coordinate information, can predict nine different QM9 properties with reasonable accuracy using a straightforward neural network.3PubMed. Machine Learning Prediction of Nine Molecular Properties Based on the SMILES Representation of the QM9 Quantum-Chemistry Dataset
The thermodynamic properties in QM9 are studied less intensively but still matter. Zero-point vibrational energy tells you how much energy a molecule retains even at absolute zero, while heat capacity describes how a molecule absorbs thermal energy. These quantities come up in chemical engineering, atmospheric chemistry, and materials science. Graph neural network research has examined how different mathematical activation functions affect prediction accuracy for both electronic and thermal properties in QM9, including dipole moment, polarizability, and specific heat capacity.4Chemical Physics. Performance analysis of activation functions in molecular property prediction using Message Passing Graph Neural Networks
QM9 as the Proving Ground for Molecular Machine Learning
Since around 2012, QM9 has served as the go-to benchmark for anyone building a new machine learning model for chemistry. If you develop a novel graph neural network, a new molecular representation, or a better training strategy, the first thing the community expects you to do is show how it performs on QM9. The dataset’s role is analogous to what MNIST was for image recognition or what the Penn Treebank was for language modeling: a shared test bed where results can be compared across labs and methods.
The progression of results on QM9 tells a story of rapid improvement. Early kernel-based methods brought prediction errors for molecular energies and HOMO/LUMO energies down toward chemical accuracy (about 1 kilocalorie per mole), and successive generations of models have continued to push those errors lower.1Journal of Cheminformatics. Dataset’s chemical diversity limits the generalizability of machine learning predictions The friendly competition that QM9 enables has driven innovation in how molecules are represented computationally. Early work used Coulomb matrices, which encode the electrostatic interactions between atom pairs. These were effective but had a notable blind spot: they cannot distinguish mirror-image molecules (enantiomers), because the pairwise distances are identical for both forms. Newer representations based on Cartesian coordinates can capture that difference.5PubMed Central. Machine Learning Classification of Chirality and Optical Rotation Using a Simple One-Hot Encoded Cartesian Coordinate Molecular Representation
More recently, the benchmark race has shifted toward architectures that exploit three-dimensional geometry. Models like SchNet, DimeNet++, Equiformer, and FAENet each take a different approach to processing spatial information, from continuous-filter convolutions to directional message passing to equivariant transformers. These architectures have been benchmarked head-to-head on QM9 to evaluate how well different strategies for encoding geometry improve prediction accuracy.6arXiv. Understanding the Capabilities of Molecular Graph Neural Networks in Materials Science Through Multimodal Learning and Physical Context Encoding
Generating New Molecules, Not Just Predicting Properties
QM9’s role has expanded beyond property prediction into generative modeling: training AI systems to invent entirely new molecules. Diffusion-based models, which are the same class of algorithm behind image generators, have been adapted to create three-dimensional molecular structures from scratch. QM9 is a standard training set for this work because its molecules are small enough that the generation task stays tractable, and the known properties provide a built-in check on whether the generated molecules are chemically sensible.7PubMed Central. Comprehensive Benchmark Study of Diffusion-Based 3D Molecular Generation Models
A typical setup divides QM9 into about 100,000 training molecules, 18,000 for validation, and 13,000 for testing. Researchers then evaluate whether a generative model can produce novel molecules that look chemically valid, have stable three-dimensional geometries, and exhibit property distributions matching the training set.8PubMed Central. Geometry-complete latent diffusion model for 3D molecule generation At least nine different state-of-the-art diffusion models have been comprehensively benchmarked on QM9 and a second, larger dataset of drug-like molecules.7PubMed Central. Comprehensive Benchmark Study of Diffusion-Based 3D Molecular Generation Models
The appeal for drug discovery researchers is clear: if you can train a model to understand the rules of molecular structure and properties from QM9, you can potentially guide it to generate molecules with specific desired characteristics. Some pipelines now use graph neural networks trained partly on QM9 to predict quantum-mechanical properties alongside pharmaceutical-relevant ones like lipophilicity and synthetic accessibility, creating a unified workflow that moves from property prediction through molecule generation to validation.9PubMed. AI-driven molecular modeling and design: from property prediction to drug generation
The Generalization Problem
Here is where QM9’s strengths become its weaknesses. A model that performs brilliantly on molecules with at most nine heavy atoms drawn from five elements does not necessarily work on bigger, more complex molecules. Drug candidates typically contain 20 to 50 or more heavy atoms and include elements like sulfur, chlorine, phosphorus, and bromine. The chemical space of QM9 is a tiny, well-manicured corner of the vast landscape that matters for real-world applications.
This gap has been documented directly. Researchers have noted that the chemical diversity of a training dataset fundamentally limits how far machine learning predictions can generalize.1Journal of Cheminformatics. Dataset’s chemical diversity limits the generalizability of machine learning predictions Training on QM9 alone does not prepare a model for the real heterogeneity of pharmaceutical chemistry. To address this, some groups have created extensions. One notable effort, QM9-extended, added roughly 20,000 molecules containing sulfur and chlorine atoms, broadening the elemental coverage to better represent drug-like chemical space.10PubMed. Exploring Deep Learning of Quantum Chemical Properties for Absorption, Distribution, Metabolism, and Excretion Predictions
Not all models struggle equally with this size gap, though. One recent graph neural network, X2-GNN, was trained exclusively on QM9 yet managed to generalize credibly to molecules with tens of heavy atoms, achieving reasonable per-atom error rates on much larger structures.11PubMed. X2-GNN: A Physical Message Passing Neural Network with Natural Generalization Ability to Large and Complex Molecules This suggests that the architecture of a model, specifically how it encodes physical interactions, can partially compensate for the limited chemical diversity of the training data. But “partially” is doing a lot of work in that sentence. Most architectures still need exposure to larger molecules to predict their properties reliably.
Questions About Data Quality
QM9’s reputation as a clean dataset is mostly deserved, but it is not without wrinkles. An unsupervised learning study that mapped the internal structure of QM9 found a distinctive pattern: the dataset has a densely packed core of “well-behaved” molecules surrounded by a broad outer region of scattered outliers. Molecules with very few atoms (fewer than 11 total, including hydrogens) or very many (more than 25) tend to land in the outlier zone, while those in the middle range cluster together neatly.12PubMed Central. Understanding the Structure of QM7b and QM9 Quantum Mechanical Datasets Using Unsupervised Learning This matters because machine learning models trained on QM9 may learn the core region well while struggling with outliers at the edges, and a researcher who is unaware of this structure might not realize that their model’s good average performance masks poor predictions for certain molecule types.
A deeper concern involves the accuracy of the underlying quantum-chemical calculations themselves. QM9’s properties were all computed with one specific density functional method. While B3LYP is widely used and generally reliable for the types of properties QM9 reports, it has known systematic biases. For instance, different density functionals can disagree on energy barriers and orbital energies, and B3LYP is known to overestimate certain quantities compared to higher-accuracy methods.13Journal of Chemical Theory and Computation. Reactive Chemistry at the Unrestricted Coupled Cluster Level: High-Throughput Calculations for Training Machine Learning Potentials This means a machine learning model trained on QM9 is learning to reproduce B3LYP’s version of reality, which is not quite the same as actual experimental reality.
Revising the Entire Dataset
The accuracy concern has motivated at least one major revision. Researchers developed an adaptive hybrid density functional called aPBE0, which adjusts its parameters for each molecule rather than using a one-size-fits-all approach. When applied to QM9, this method produced systematically different results: stronger covalent bonding, larger HOMO-LUMO gaps, more localized electron densities, and larger dipole moments compared to the original B3LYP calculations. The resulting revision, called revQM9, is estimated to be vastly superior in quality to the original.14arXiv. Adaptive hybrid density functionals
The existence of revQM9 raises a practical question for anyone using QM9 for benchmarking: which version should you train on? If you use the original, your results are directly comparable to years of published work, but the underlying property values may be systematically off. If you use the revision, your training data is arguably more accurate, but your results are not directly comparable to the existing literature. Most published work as of now still uses the original QM9, partly out of inertia and partly because the benchmark’s value lies in comparability across studies.
Bigger Datasets and Where QM9 Fits Now
QM9 is no longer the only game in town for molecular machine learning. Larger and more diverse datasets have appeared to address its limitations. The OGB-LSC PCQM4Mv2 dataset, drawn from PubChemQC, contains over 3.7 million molecules with calculated HOMO-LUMO gaps and three-dimensional conformations for the majority of them.15Nature Communications. Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing That is roughly 28 times the size of QM9, with substantially greater molecular diversity.
Other efforts have pursued higher accuracy rather than larger size. The ANI-1ccx dataset, for example, contains about 500,000 data points computed with coupled-cluster methods, which are considerably more accurate than the density functional theory used for QM9. The companion ANI-1x dataset provides around 5 million density-functional calculations. Together, these datasets required approximately 14 million CPU core-hours to generate, illustrating the enormous computational cost of building high-quality quantum-chemical reference data.16PubMed Central. The ANI-1ccx and ANI-1x data sets, coupled-cluster and density functional theory properties for molecules.
Despite these alternatives, QM9 retains a distinct role. Its small size and simplicity make it fast to work with, which matters when you are iterating on model architecture and need quick feedback. Its well-characterized structure means you can diagnose model failures more easily than with a messier, larger dataset. And the mountain of published results on QM9 provides an unmatched reference point. Newer datasets tend to complement QM9 rather than replace it: researchers often validate a new method on QM9 first for comparability, then test on a larger dataset for real-world relevance.
Hybrid Quantum-Classical Approaches and Drug Discovery Pipelines
One frontier where QM9 keeps appearing is the intersection of quantum computing and molecular machine learning. Hybrid quantum-classical frameworks are being explored for drug discovery, using quantum-inspired algorithms to screen molecular candidates more efficiently than purely classical approaches.17Procedia Computer Science. A Hybrid Quantum-Classical Machine Learning Framework for Accelerated Drug Discovery QM9 serves as a testing ground for these methods because its molecules are simple enough to be tractable on near-term quantum hardware, where qubit counts and coherence times are still limited.
The pipeline connecting QM9-trained models to actual pharmaceutical work is longer than it might sound. A model that predicts HOMO-LUMO gaps for nine-atom molecules is still several steps removed from predicting whether a candidate drug will be absorbed by the gut, metabolized by the liver, or bind to a target protein. Bridging that gap requires not just scaling up the training data (as with QM9-extended) but also training on task-specific endpoints that QM9 does not contain, like solubility, protein binding affinity, and toxicity. The most promising current approaches use QM9 and similar datasets to pre-train models on fundamental physics, then fine-tune on smaller, more application-specific datasets. The quantum-mechanical understanding learned from QM9 provides a foundation that helps models make better predictions even on unfamiliar molecules and unfamiliar properties.
What Researchers Still Get Wrong About QM9
A few misconceptions circulate in the broader machine learning community about what QM9 results actually mean. The most common is treating low error on QM9 as proof that a model “understands” chemistry. QM9 is small, clean, and narrowly scoped. Achieving near-perfect prediction accuracy on it demonstrates that your model can interpolate within a well-defined chemical space, which is useful but quite different from understanding chemical behavior across the periodic table.
A related mistake is assuming that QM9 properties are ground truth. They are computed properties from a specific level of theory, not experimental measurements. The original B3LYP calculations contain systematic biases, and the revQM9 revision showed that even switching to a modestly better functional changes the numbers across the board.14arXiv. Adaptive hybrid density functionals A model that perfectly reproduces QM9’s HOMO values has perfectly learned B3LYP’s approximation of HOMO values, which is one step removed from physical reality. For many applications that is fine, but it is worth keeping the distinction in mind, especially when comparing model predictions against experimental data.
Finally, QM9’s outlier structure means that aggregate error metrics can be misleading. If most of your test molecules fall in the well-clustered core region, your average error will look great even if your model completely fails on the scattered outlier molecules at the periphery.12PubMed Central. Understanding the Structure of QM7b and QM9 Quantum Mechanical Datasets Using Unsupervised Learning Researchers who report only mean absolute error on QM9 without examining performance across different molecular subsets may be painting a rosier picture than reality warrants.