Energy-Density Baselines

A CIDER kernel combines a learned scalar function with baseline energy-density models. For the original DFTKernel interface, Python callables evaluate

\[e(\mathbf X)=a(\mathbf X)+m(\mathbf X)f_\mathrm{ML}(\mathbf X)\]

and return the corresponding derivatives with respect to the raw features. In the above equation, \(a(\mathbf X)\) and \(m(\mathbf X)\) are additive and multiplicative baselines, respectively, and \(f(\mathbf X)\) is the function learned via Gaussian process regression. zero_xc and one_xc provide constant additive or multiplicative terms; the exchange and correlation helpers provide LDA/GGA factors. These callables are selected during model construction and serialized into the mapped kernel.

DFTKernel2 stores libxc identifiers for \(a\) and \(m\). MappedDFTKernel2 evaluates their energy, density, gradient, and kinetic-energy-density derivatives through get_libxc_baseline. This is the baseline path used by the separate exchange and correlation kernels in CIDER26XC.