Séminaire LCPMR | Damien Bastier "FCInet : A classification neural network to describe the configuration interaction space"

FCInet : a classification neural network to describe the configuration interaction space

Bastien Casier

In this study, we present an iterative machine learning classification algorithm to smartly select the most important Slater determinants used in the expansion of the electronic wave-function. The learnable information is encoded through a binary representation of the spin-orbital populations of each Slater determinant, while the learning is based on the minimization of the binary cross-entropy.
Today, this method presents very promising results for small molecules like CO, N2, H2O, NH3 and C2H6. The FCI accuracy have been obtained along the dissociation curve of diatomic systems by only selecting a small fraction of the Hilbert’s space. This primary study is a proof of principle of the binary classification of the Slater determinants for the selective CI methods.
Our end goal would be to realize a global learning on the full dissociation curves to define the first interaction potentials at the FCI level.