A scalable diagonalization framework for tensor-product bitstring selected configuration interaction

Area of Science:

• Quantum Chemistry
• Computational Physics
• High-Performance Computing

Background:

• Selected configuration interaction (SCI) methods are crucial for accurately describing strongly correlated electronic systems.
• Existing SCI implementations face scalability limitations due to memory-intensive replication of configuration interaction (CI) vectors.

Purpose of the Study:

• To develop a fully distributed diagonalization framework for selected determinant spaces to address SCI scalability bottlenecks.
• To introduce and validate the tensor-product bitstring SCI (TBSCI) method.

Main Methods:

• A novel tensor-product bitstring (TPB) representation organizes determinants using selected alpha- and beta-bitstrings.
• An efficient TBSCI eigensolver was developed, incorporating a bitstring-based Hamiltonian evaluation and MPI communication strategies.
• Large-scale full configuration interaction (FCI) benchmarks were used to test the distributed diagonalization.

Main Results:

• The TBSCI eigensolver significantly reduces wall time for distributed diagonalization of extremely large determinant spaces (2.6 × 10^12 determinants).
• Scalability was demonstrated on the Fugaku supercomputer, utilizing 54,000 nodes (>2.5 million cores).
• The TPB representation shows intrinsic compactness, enabling wavefunctions close to the FCI limit with a reduced number of determinants.

Conclusions:

• TBSCI establishes a new scalable SCI methodology, overcoming previous computational limitations.
• The TPB representation offers a compact and efficient way to represent wavefunctions in large electronic systems.
• This work paves the way for accurate quantum chemical calculations on previously intractable strongly correlated systems.