QuantumCanvas: A Multimodal Benchmark for Learning Two-Body Quantum Interactions

QuantumCanvas is a large-scale multimodal benchmark that treats two-body quantum systems as a minimal, systematically enumerable unit of interatomic interaction. It spans 2,850 element–element pairs across 75 elements, each annotated with 20 electronic, thermodynamic, and geometric property targets and paired with a ten-channel, coordinate-free orbital-image tensor derived from l- and m-resolved orbital densities, angular field transforms (Gramian angular field, Markov transition field), co-occupancy maps, and charge-density projections. This archive contains the processed dataset dataset_combined.npz. Arrays: images (2850, 10, 32, 32), geometries (2850, 2, 4), elements (2850, 2), labels (2850,) (per-pair dict of 20 targets), metadata (2850,), pair_names (2850,). See DATA_README.md in this record for the full schema. All labels are computed at the SCC-DFTB level (DFTB+, PTBP parameter set). The dataset can be regenerated from the DFTB+ outputs by running python build_dataset.py from the code repository. Code & generation pipeline: https://github.com/KurbanIntelligenceLab/QuantumCanvas (MIT License).Paper: "QuantumCanvas: A Multimodal Benchmark for Learning Two-Body Quantum Interactions"