CiderPress: Machine-Learned Exchange-Correlation Functionals

CiderPress implements the CIDER framework for constructing and evaluating machine-learned density functionals. The code serves several purposes:

  • Specify the physically constrained electronic descriptors used as model inputs for CIDER functionals, along with the numerical settings for evaluating these features.

  • Train, evaluate, store, and load Gaussian process regression models representing CIDER functionals. Tools are also included to map Gaussian processes to more efficient inference-time models.

  • Through interfaces to existing DFT backends (PySCF [1] and GPAW [2]), compute the electronic input descriptors, evaluate the XC energy and potential, and perform full self-consistent field calculations with CIDER functionals.

CIDER stands for Compressed scale-Invariant DEnsity Representation. The name originally described the scale-invariant density features used to learn exchange. The framework now also includes smoothed density-matrix features, full exchange-correlation models, and molecular and periodic numerical implementations.

Getting started

The calculation guides state which combinations of models, settings, and DFT backends are supported. In brief, PySCF is the molecular all-electron backend, while classic GPAW provides the periodic plane-wave PAW implementation. The packaged functional families compute different parts of the XC energy. CIDER23X and CIDER24X use an explicit surrogate-hybrid composition; CIDER26XC uses the full-XC initialization described in the model guide.

Getting started

Running calculations

The CIDER framework

Model and descriptor workflows

API and implementation reference

Reference