UnMixNet: Physics-Closed Graph Solver Boosts Machine Olfaction Accuracy

Machine olfaction—inferring gas compositions from sensor arrays—is an underconstrained inverse problem. Sensor responses are low-dimensional, delayed, and entangled by interacting chemical transport, surface adsorption, and transduction. Traditional neural networks often suffer from physics closure misspecification: they fit sensor traces rather than underlying physical processes. This leads to poor generalization, especially on unseen mixtures or dynamic transitions.

To address this, Yue Shi and colleagues propose UnMixNet, a physics-closed graph neural solver. It formulates gas unmixing as a multi-physics-constrained inverse problem governed by Maxwell-Stefan cross-diffusion PDEs, competitive adsorption ODEs, and nonlinear sensor transduction ODEs. UnMixNet discretizes these equations on spatial graphs, computing multicomponent flux on each edge. This enables local, differentiable, and flux-conservative inference. Evaluated on SmellNet, it outperformed baselines in single-odor, seen-mixture, and unseen-mixture tasks. External validation on the UCI Dynamic Gas Mixtures dataset showed inferred concentration processes matched ground truth set points during dynamic transitions. Process-consistency diagnostics confirmed the model learns transferable physical fingerprints satisfying transport, conservation, adsorption, and readout closure.