Sanibel Symposium

Tensor network theory provides a versatile, flexible, and often highly efficient route to multi-linear compression of tensors in quantum many-body theory and beyond, for example, in generic multivariate data analytics. In particular, in molecular electronic structure theory, the configuration interaction and coupled cluster tensors have successfully been factored using specific tensor decompositions, which results in a reduction of the computational cost of the corresponding approximate methods. To facilitate adoption of different tensor network approximations (tensor factorizations) in HPC applications, we started an open source C++ library ExaTN. Our library provides a generic tensor network builder as well as some algorithmic primitives required for the optimization of the tensor network factors, such that one can construct a given tensor network and obtain explicit tensor equations necessary for the optimization of the constituting tensor factors. The obtained tensor equations can then be scheduled for execution via a suitable numerical tensor algebra backend, either on shared or on distributed memory platforms. We hope that our effort at this early stage of development could provide a future reusable layer for many HPC applications interested in tensor network approximations. In particular, we are looking forward into the implementation of the tree tensor networks and multiscale entanglement renormalization ansatz based on ExaTN.

This work has been supported by the Laboratory Directed Research and Development program at Oak Ridge National Laboratory (grant #9463). This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725.