Double Precision Is Not Needed for Many-Body Calculations: Emergent Conventional Wisdom

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Traditional electronic structure programs utilize double-precision floating-point representation, leading to substantial memory and computation requirements for large-scale simulations.
• The memory footprint and input/output bottlenecks in many-body theories and large-scale calculations can be significantly reduced by using smaller data representations.

Purpose of the Study:

• To investigate the feasibility and accuracy of employing single-precision floating-point representation in coupled-cluster (CC) and equation-of-motion coupled-cluster (EOM-CC) methods.
• To assess the impact of single-precision calculations on the accuracy of computed energies, gradients, excited states, and molecular properties.
• To develop and evaluate an implementation that allows for significant computational savings while maintaining high accuracy.

Main Methods:

• Implementation of single and double excitation coupled-cluster (CC) and equation-of-motion coupled-cluster (EOM-CC) methods using single-precision floating-point arithmetic.
• Exploration of both standard implementations and those incorporating Cholesky decomposition or resolution-of-the-identity (RI) approximations for electron-repulsion integrals.
• Numerical validation through extensive testing to quantify accuracy loss compared to double-precision calculations.

Main Results:

• Single-precision implementation of CC and EOM-CC methods results in a factor of 2 reduction in data size and computation time.
• Numerical tests demonstrate that the loss of accuracy in correlated calculations using single precision is insignificant.
• The developed implementation allows for pure single-precision computations of energies, analytic gradients, excited states, and molecular properties with high fidelity.

Conclusions:

• Single-precision coupled-cluster calculations are viable and offer substantial computational advantages without compromising accuracy for key molecular properties.
• The implementation provides a pathway to fully recover double-precision accuracy through optional cleanup iterations, balancing performance and precision.
• This approach significantly alleviates memory and I/O bottlenecks, enhancing the scalability of electronic structure calculations on modern hardware.