UMN: Universe Multi Neural Network for Olfactory Mixture Similarity Prediction via Dierentiable N-body Simulation

Predicting the perceptual similarity of olfactory mixtures from their molecular composition remains an open problem in computational olfaction. Existing approaches either rely on hand-crafted physicochemical descriptors with manual feature selection or employ graph neural networks that operate on individual molecules without modeling inter-molecular interactions at the mixture level. In this work, we propose Universe Multi Neural network (UMN), a physics-simulation-based architecture that maps molecular fingerprints to celestial bodies with learnable mass, position, and velocity, performs differentiable N-body gravitational simulation, and extracts orbital stability features as mixture-level representations. We evaluate UMN on the Snitz et al. (2013) mixture similarity dataset (360 pairs) using 5-seed × 5-fold cross-validation. UMN achieves a Pearson correlation of r = 0.780 (std = 0.039) without any manual feature engineering, compared to r >= 0.49 for a non-optimized descriptor-based model. Through systematic experiments across 9 versions (v17-v25), we demonstrate that (1) the physics engine architecture is near-optimal in its original form, (2) multi-restart training is the most effective optimization strategy, and (3) contrastive learning, physics-based loss functions, and architectural modifications all degrade performance on this small-data regime.