QuantumCanvas: a multimodal benchmark for learning two-body quantum interactions

Abstract Most molecular and materials machine-learning models fit correlations across whole molecules or crystals rather than learning the quantum interactions between atomic pairs. Yet bonding, charge redistribution, orbital hybridization, and electronic coupling all emerge from these two-body interactions. We introduce QuantumCanvas, a multimodal benchmark that treats the two-body quantum system as the minimal, exhaustively enumerable unit of interatomic interaction. It covers 2,850 element–element pairs, each at a single optimized geometry, and evaluates 17 benchmark target quantities spanning electronic, thermodynamic, dipole, and charge-derived properties; algebraically derived quantities and redundant diagnostic charge rows are identified explicitly, and label availability is reported per target. Each pair is also represented by ten-channel images of orbital populations and charge- and dipole-derived fields that encode angular and electrostatic structure without explicit atomic coordinates. Benchmarking graph, vision, and fusion architectures on element-pair-disjoint splits reveals modality-specific inductive biases: graph encoders achieve the lowest MAE on most reported targets, while late fusion gives the lowest MAE for the reported Mermin free-energy label. Controls on the energy gap and dipole magnitude show that destroying the spatial layout of the images does not degrade accuracy and that a model fed the generating scalars directly outperforms both image variants: the rendering is an alternative encoding of the same scalars, not an independent signal. Pretraining on QuantumCanvas lowers mean test error in 11 of 16 encoder–target comparisons across QM9, MD17, and CrysMTM. QuantumCanvas thus provides a controlled, physically grounded testbed for studying which signals each modality captures, how they combine, and how they transfer across molecular, dynamical, and crystalline regimes.