Optimal Orbital Selection for Full Configuration Interaction (OptOrbFCI): Pursuing the Basis Set Limit under a Budget.

Yingzhou Li, Jianfeng LuJournal of Chemical Theory and Computation 2020 · OpenReview

Abstract

The Full Configuration Interaction (FCI) method represents the exact solution to the electronic Schrödinger equation within a chosen basis set, yet its exponential scaling limits its application to small systems or truncated basis sets. We present OptOrbFCI, an algorithmic framework that optimizes the selection of a small number of orbitals from a much larger basis set to minimize the ground-state energy. By treating orbital selection as a constrained optimization problem on the Stiefel manifold and employing alternating minimization with Barzilai-Borwein steps, OptOrbFCI effectively "compresses" the physics of large basis sets into a manageable computational budget. Our results on $H_2O$, $C_2$, and $N_2$ demonstrate that OptOrbFCI consistently outperforms traditional CASSCF methods and provides a more accurate description of molecular dissociation.

1. Introduction

In the field of quantum chemistry, the Full Configuration Interaction (FCI) method is often referred to as the "gold standard" for its ability to provide an exact solution for the electronic Schrödinger equation within a defined basis set. However, the computational cost of FCI scales exponentially with the number of orbitals, creating a significant bottleneck. Researchers are frequently forced to choose between a small, computationally feasible basis set (e.g., cc-pVDZ) that lacks physical accuracy, or a large, accurate basis set (e.g., cc-pV5Z) that is computationally inaccessible.

The OptOrbFCI framework addresses this dilemma by asking a fundamental question: Given a computational budget of $N$ orbitals, which $N$ linear combinations of a much larger set of $M$ orbitals will yield the lowest possible ground-state energy? By answering this, OptOrbFCI allows for the "compression" of large-basis-set accuracy into a small-basis-set cost.

2. Methodology

OptOrbFCI formulates orbital selection as a formal optimization problem. The objective is to minimize the ground-state energy $E$ over two coupled variables: the many-body wave function $|\Phi\rangle$ and a partial unitary matrix $U \in \mathbb{R}^{M \times N}$ that defines the orbital rotation and selection.

2.1 Alternating Minimization

The complexity of the problem necessitates an alternating minimization approach, toggling between two primary steps:

  1. Wave Function Optimization (Fixed Orbitals): With the orbital set fixed, the algorithm solves the FCI problem. To maintain efficiency, the authors utilize the Coordinate Descent FCI (CDFCI) solver, which is well-suited for identifying the most significant determinants in the wave function.
  2. Orbital Optimization (Fixed Wave Function): With the wave function fixed, the algorithm optimizes the orbital rotation matrix $U$.

2.2 Optimization on the Stiefel Manifold

A key innovation of OptOrbFCI is its treatment of the orbital rotation matrix $U$ as an element of the Stiefel manifold—the space of matrices with orthonormal columns. Unlike traditional Complete Active Space Self-Consistent Field (CASSCF) methods, which often rely on local second-order approximations like Newton-Raphson, OptOrbFCI employs a projection method with alternating Barzilai-Borwein (BB) stepsizes. This first-order approach is more robust in navigating the complex "orbital landscape."

Furthermore, to prevent the algorithm from becoming trapped in local minima—a common issue in CASSCF—the authors introduce small random perturbations during the orbital optimization phase. This "shaking" mechanism helps the system explore the energy surface more thoroughly to find a more global minimum.

3. Results and Discussion

The performance of OptOrbFCI was evaluated using standard molecular benchmarks, including water ($H_2O$), the carbon dimer ($C_2$), and the nitrogen dimer ($N_2$), using Dunning’s correlation-consistent basis sets (cc-pVXZ).

3.1 Basis Set Compression in $H_2O$

In tests on the water molecule, OptOrbFCI demonstrated the ability to achieve lower ground-state energies than standard CASSCF implementations for the same number of active orbitals. This indicates that the BB-step optimization and random perturbations are more effective at finding the optimal orbital subspace than traditional local search methods.

3.2 Nitrogen Dimer Dissociation

The most striking result involves the dissociation energy of $N_2$. A standard FCI calculation using a cc-pVDZ basis set (28 orbitals) yields a dissociation energy of 0.3163 Ha. By using OptOrbFCI to select 28 "optimal" orbitals from a much larger cc-pVQZ set (110 orbitals), the researchers achieved a dissociation energy of 0.3522 Ha. This "compressed" calculation significantly improves accuracy, bringing the result closer to the basis set limit while maintaining the computational cost associated with only 28 orbitals.

4. Conclusion and Future Directions

OptOrbFCI provides a powerful framework for extending the reach of high-accuracy quantum chemistry methods. By optimizing the orbital basis directly, it allows researchers to bypass the traditional trade-off between basis set size and computational feasibility.

While the current results are promising, several avenues for future work remain. The current implementation does not yet account for molecular symmetry or "frozen core" approximations, which results in dense integrals and increased memory overhead. Additionally, the method is currently focused on ground-state calculations; extending the framework to excited states is essential for broader chemical applications. Finally, the authors note that the orbital optimization step is highly parallelizable and could benefit significantly from GPU acceleration, further enhancing the efficiency of the "basis set compression" process.

Summary updated

BibTeX
@article{li2020optimal,
  author = {Yingzhou Li and Jianfeng Lu},
  title = {{Optimal Orbital Selection for Full Configuration Interaction (OptOrbFCI): Pursuing the Basis Set Limit under a Budget.}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2020},
  url = {https://openreview.net/forum?id=nyFUlVXexL},
}
Figure