Quantum evolution kernel for molecular machine learning on programmable neutral-atom arrays

Neutral-atom quantum processors can embed molecular graphs natively through the geometry of atom arrays, which makes them a natural platform for graph-based quantum kernels. Existing quantum evolution kernels, however, encode limited chemical information and have been applied to molecular classification but not to regression. Here we introduce ChemQEK, a quantum evolution kernel framework for molecular property prediction on analog neutral-atom quantum processors. ChemQEK encodes atom-specific descriptors through local addressability and atom-pair London dispersion coefficients ($C^{ij}_6$) into the parameters of the Rydberg Hamiltonian, and it combines global, local and pairwise observables of the evolved state within a single kernel. We further introduce kernel pooling strategies that combine information across evolution times and addressability schemes. Using noiseless emulation, we evaluate ChemQEK with kernel-based regression models on the ESOL, FreeSolv, Jean-Claude Bradley Open Melting Point, and QM9 datasets. Pooling over both evolution times and addressability schemes outperforms pooling over either alone and yields test $R^2$ values of 0.90, 0.84, and 0.69 for ESOL, FreeSolv, and melting point, respectively. On QM9, ChemQEK outperforms the quantum evolution kernel and generalized-distance quantum-correlation frameworks, each of which relies on a single observable type, for all eight properties considered. The QM9 experiments also reveal property dependent trade-offs between predictive performance and the number of quantum feature configurations, with up to $\sim$79\% fewer measurements retaining competitive performance for several properties. These results highlight the importance of chemically informed encoding, complementary observables, and multidimensional kernel pooling for quantum kernel-based molecular property prediction.