Checklist APOAI 2026 Contest · Task 3
Molecular Energy and Dipole Moment Prediction
Predict the energy and dipole-moment magnitude of QM9 molecules from 3D atomic coordinates, training on only 200 molecules.
The task
Computational chemistry trades accuracy against cost; the contestant builds a model that predicts a molecule's energy (atomic units) and dipole-moment magnitude (Debye) from its atomic coordinates. Both targets are invariant to translation and rotation.
Coordinates are sampled from QM9. Each molecule is an XYZ file (atom count, name, then element symbol and Cartesian coordinates in Å); molecules are neutral singlets with at most 50 atoms from H, C, N, O and F. The training set has 200 molecules (train/molecule_N/coordinate.xyz and train/training_ref.csv); test groups A and B have 300 molecules each and are only available on the evaluation machine. Optional feature-engineering ideas (atomic properties, distance-based invariants, centred coordinates, one-hot atom types) are suggested.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
- 200 training XYZ files with
training_ref.csv(molecule #, energy, dipole moment); test XYZ files for groups A and B on the evaluation machine. - You submit
submission_A.csvandsubmission_B.csv(columns: molecule #, energy, dipole moment; 300 rows each) packed intosubmission.zip.- Scoring
- Per group: Score_energy = max[0, 0.50 × (1 − 5 × MRE_energy)] with MRE the mean relative error; Score_dipole = max[0, 0.50 × (1 − MAE_dipole / 5)]; total = sum (maximum 1). Baseline on B: 0.7729; scientific committee reference: 0.8758.
- Rules
- Pretrained model weights are not allowed, including weights pretrained on ImageNet or any other external dataset.
- External large language model APIs (e.g. GPT, Claude) may not be used for prediction, feature generation, data labelling or model ensembling.
- Single training run of the notebook ≤ 20 minutes on CPU.
- No connectivity or bond information is given; everything must be inferred from coordinates.
- Evaluated in the 'ioai3.6' image.
- Format
- APOAI 2026 contest (first edition): a six-hour examination on Saturday 27 June 2026, taken online on the Bohrium platform with local invigilated sites; the four tasks were republished on Bohrium for practice from 15 June to 31 July 2026 (platform time UTC+8).