Alchemical derivatives provide a systematic description of molecular energy responses to continuous changes in nuclear charge, enabling efficient exploration of chemical space within the framework of quantum alchemy. In this work, first- and second-order alchemical derivatives are computed using alchemical perturbation density functional theory and differentiable quantum chemistry frameworks. These derivatives are employed as noise-free molecular representations for kernel ridge regression (KRR). A linear alchemical kernel capturing additive atomic contributions is constructed and combined with nonlinear kernels via kernel addition to enhance model expressivity. This combined kernel approach leads to reduced mean absolute errors and improved data efficiency compared to nonlinear kernels alone. Overall, our results will demonstrate that alchemical derivative-based representations serve as physically motivated and effective fingerprints for kernel-based molecular property prediction.
Sana Qureshi