Analytical Nuclear Gradients for State-Averaged Configuration Interaction Singles Variants: Application to Conical Intersections

This paper derives analytical nuclear gradients for state-averaged orbital-optimized configuration interaction singles (SACIS) and its spin-projected variant (SAECIS), demonstrating that these low-cost methods accurately reproduce conical intersection geometries and topologies by effectively capturing static correlation through orbital relaxation, thereby offering a reliable black-box alternative to high-level theories at mean-field computational cost.

The Gist

The Big Picture: Finding the "Secret Pass" in a Mountain Range

Imagine you are trying to navigate a massive, foggy mountain range. This landscape represents the Potential Energy Surface of a molecule. The height of the land is the energy of the molecule; the lower the valley, the more stable the molecule is.

Usually, molecules travel along smooth paths from one valley to another. But sometimes, two different valleys (representing two different electronic states of the molecule) crash into each other at a single point. This point is called a Conical Intersection (CX).

Think of a Conical Intersection as a secret mountain pass. If a molecule hits this pass, it can instantly switch from one "state" (like being excited by sunlight) to another (like returning to a calm state) without losing energy as heat. This is how plants photosynthesize and how our eyes see light. It happens incredibly fast—faster than a blink.

The Problem: The Old Maps Were Wrong

To find these secret passes, scientists use computer programs to draw maps of the mountains. However, the old, cheap maps (methods like standard CIS or TDDFT) had a fatal flaw: they were too simple.

• The Flaw: These old methods assumed the ground was flat and smooth. They couldn't see the "cliff" where the two valleys met. When they tried to draw the map near the pass, the lines would just stop or jump discontinuously. It was like trying to draw a map of a bridge using only a ruler; you miss the curve entirely.
• The Expensive Fix: There are "Gold Standard" maps (like SA-CASSCF or MRCI) that are incredibly accurate. They show every rock and tree. But they are so computationally expensive that they are like hiring a team of 100 surveyors to walk every inch of the mountain. You can't use them for huge, complex molecules (like proteins or DNA) because the calculation would take longer than the age of the universe.

The Solution: A New, Smart Compass

The author, Takashi Tsuchimochi, invented a new type of compass and a new way to draw maps. He calls it SACIS (State-Averaged Configuration Interaction Singles) and its slightly more complex cousin, SAECIS.

Here is how they work, using simple analogies:

1. The "State-Averaged" Approach: Looking at Two Hikers at Once

Old methods tried to map the "Ground State" (Hiker A) perfectly, then map the "Excited State" (Hiker B) perfectly. But near the secret pass, Hiker A and Hiker B are practically the same person. If you treat them separately, you get confused.

SACIS says: "Let's stop treating them as separate people. Let's optimize the map for both hikers at the same time." By averaging their needs, the method finds a path that works for both, ensuring the map is smooth right where the two valleys meet.

2. The "Orbital Relaxation": Flexible Shoes

Imagine the old methods were like wearing stiff, heavy boots. They were optimized for walking on flat ground (the ground state). When the hiker tried to climb the steep, twisted cliff of the conical intersection, the boots didn't bend, and the hiker fell.

SACIS gives the hiker flexible, custom-molded shoes. As the molecule twists and turns near the intersection, the "shoes" (the mathematical orbitals) reshape themselves instantly to fit the terrain. This allows the method to "feel" the static correlation (the complex, tangled nature of the electrons) that the stiff boots missed.

3. The "Spin Projection": The Magic Filter

Sometimes, the flexible shoes get a little wobbly. The SAECIS method adds a "Magic Filter" (Spin Projection). It takes the wobbly, broken-symmetry shape and forces it to snap back into a perfect, symmetrical shape.

• The Result: The paper found that for most mountain passes, the flexible shoes (SACIS) were enough. The Magic Filter (SAECIS) didn't change the shape of the pass much, but it might help if you are climbing a very steep, weird cliff (involving complex double-excitations).

4. The "Null Space" Problem: Removing the Ghosts

This is the most technical part, but here is the analogy:
When the computer tries to solve the math for these flexible shoes, it sometimes gets confused by "ghosts"—mathematical solutions that look like answers but are actually nonsense (like dividing by zero). If you don't remove these ghosts, your map becomes jagged and useless.

The author developed a special Ghost-B-Gone technique. He explicitly identifies these nonsense solutions and projects them out of the equation. This ensures the final map is smooth, stable, and physically real.

The Results: A Cheap Map That Works Like a Gold Standard

The author tested these new methods on 12 different "mountain passes" (molecules like ethylene, butadiene, and stilbene).

• The Comparison: He compared his new maps (SACIS/SAECIS) against the "Gold Standard" (MRCI) and the "Old Cheap Maps" (CIS).
• The Finding:
  • The Old Cheap Maps failed completely; they couldn't even see the pass.
  • The Gold Standard was perfect but slow.
  • SACIS and SAECIS were the sweet spot. They produced maps that were almost identical to the Gold Standard (within 0.1 Angstroms, which is smaller than an atom!) but ran much faster.

The Conclusion: Which Tool Should You Use?

The paper concludes with a practical recommendation:

1. For most situations: Use SACIS. It's the "Swiss Army Knife." It's fast, cheap, and gives you a qualitatively correct picture of the secret pass. You don't need the extra complexity of the Magic Filter (Spin Projection) for 90% of cases.
2. For special, tricky cliffs: Use SAECIS. If you are dealing with a very high-energy state that has a weird "double-excitation" character (like a molecule that is excited in two ways at once), the Magic Filter helps. But be warned: it costs more computing power.

Summary in One Sentence

The author created a fast, efficient, and mathematically stable way to map the "secret mountain passes" where molecules change states, proving that you don't need a supercomputer to get a qualitatively correct map if you use the right flexible shoes and remove the mathematical ghosts.

Technical Summary

1. Problem Statement

Conical intersections (CXs) are critical features in molecular potential energy surfaces (PES) that facilitate ultrafast non-radiative transitions between electronic states. Accurately modeling CXs requires a balanced treatment of multiple electronic states and the inclusion of static correlation to handle near-degeneracies.

• Limitations of Existing Methods:
  • Multireference Methods (e.g., SA-CASSCF): While rigorous, they suffer from high computational costs and sensitivity to active space selection, limiting their application to large systems.
  • Single-Reference Methods (e.g., CIS, TDDFT): These are computationally efficient but intrinsically fail to describe CXs because they lack static correlation and suffer from ground-state orbital bias. They often produce discontinuous PESs or fail to locate the intersection entirely.
  • Spin-Flip (SF) Methods: These improve CX descriptions but still rely on specific reference states and may not offer a fully variational framework for orbital optimization.
• The Gap: There is a need for a computationally efficient (mean-field cost), black-box method that can qualitatively describe CXs and, crucially, provide analytical nuclear gradients to enable geometry optimization and Minimum-Energy Conical Intersection (MECX) searches.

2. Methodology

The authors derive analytical nuclear gradients for two state-averaged orbital-optimized methods:

1. SACIS (State-Averaged Configuration Interaction Singles): An orbital-optimized CIS method where orbitals are optimized for an average of $n$ states.
2. SAECIS (State-Averaged Extended CIS): A spin-projected extension of SACIS that applies a spin-projection operator to broken-symmetry determinants to recover spin symmetry and access configurations with partial double-excitation character.

Key Theoretical Developments:

• Lagrangian Formulation: The authors employ a Lagrangian approach to derive the gradients. The energy functional is minimized with respect to orbital rotations ($o\lambda$) and CI coefficients ($c\lambda$).
• Handling Redundancy (Null Space): A major technical hurdle identified is that the electronic Hessian in state-averaged CI spaces is singular due to redundant parametrization. This leads to non-unique solutions in the Coupled-Perturbed (CP) equations, contaminating the Z-vector and resulting in unphysical gradients.
  • Solution: The authors introduce an explicit projection procedure to remove null-space contributions from the Z-vector solution. They construct a projector based on the CI coefficient vectors of the states to ensure numerical stability and physically meaningful gradients.
• Spin Projection: For SAECIS, the formulation accounts for the non-unitary nature of the spin-projection operator, requiring a generalized projection operator to maintain orthogonality between states during orbital perturbations.

3. Key Contributions

• Derivation of Analytical Gradients: This is the first derivation of analytical nuclear gradients for SACIS and SAECIS, enabling efficient geometry optimization and MECX searches.
• Null-Space Projection: The development of a robust algorithm to explicitly eliminate null-space components in the coupled-perturbed equations, ensuring the stability of gradient calculations for state-averaged methods.
• Black-Box Applicability: The methods are designed to be "black-box" (requiring minimal user intervention regarding active spaces) while maintaining a computational cost comparable to standard mean-field methods (CIS/TDDFT).

4. Results and Performance

A. Ethylene (Twisted-Pyramidalized) Benchmark

• Topology: Conventional CIS and ECIS failed to reproduce the correct conical intersection topology (showing discontinuities or no intersection).
• SACIS/SAECIS Success: Both SACIS and SAECIS qualitatively reproduced the correct CX topology.
  • Mechanism: SACIS achieves this through variational orbital relaxation. By optimizing orbitals for the state-averaged energy, the method induces spatial symmetry breaking and charge-transfer-like localization. This mimics the static correlation usually requiring multi-reference treatments, effectively recovering the degeneracy at the intersection.
  • Spin Projection: SAECIS yielded results very similar to SACIS for the ethylene CX, suggesting that spin projection is not strictly necessary for the qualitative description of this specific intersection, though it offers flexibility for higher excited states.

B. MECX Benchmark (12 Systems)
The authors tested SACIS and SAECIS on 12 MECXs (including ethylene, butadiene, styrene, PSB3, etc.) against high-level references (MRCI, SF-TDDFT, SA-CASSCF).

• Geometric Accuracy: Both methods achieved Mean RMSDs below 0.1 Å relative to the reference methods.
  • SACIS and SAECIS showed comparable structural accuracy.
  • SACIS geometries were often closer to MRCI or SF-TDDFT depending on the system, rather than strictly following the minimal active-space SA-CASSCF picture.
• Energetics:
  • Vertical excitation energies at the Franck-Condon point showed larger errors (MAE ~0.38–0.73 eV), consistent with the lack of dynamic correlation in CIS-based methods.
  • Adiabatic MECX energies were reproduced more consistently (MAE ~0.31–0.46 eV), indicating that the balanced orbital relaxation is particularly effective in the near-degenerate region.
• Specific Cases:
  • SAECIS Advantage: SAECIS showed an advantage in describing the $S_2$ state of ethylene (which has double-excitation character), which is inaccessible to pure single-excitation SACIS.
  • Cost-Benefit: For general CX optimization, SACIS is preferred due to lower computational overhead and faster convergence, as spin projection in SAECIS can deteriorate the conditioning of the equations without significantly improving MECX geometry accuracy.

5. Significance and Conclusion

• Efficiency vs. Accuracy: The paper demonstrates that SACIS and SAECIS provide a qualitatively reliable description of conical intersections at a mean-field computational cost. This bridges the gap between expensive multireference methods and inaccurate single-reference methods.
• Orbital Relaxation is Key: The study highlights that state-averaged orbital optimization is the primary driver for recovering static correlation and correct CX topology in CIS-based frameworks, often more so than spin projection for ground-state crossings.
• Practical Utility: The availability of analytical gradients makes these methods suitable for automated geometry optimization and dynamics simulations in large molecular systems where multireference methods are infeasible.
• Future Outlook: While currently qualitative, the framework serves as a foundation for adding dynamic correlation (e.g., perturbation theory) to achieve quantitative accuracy.

In summary, this work establishes SACIS as a highly efficient, black-box tool for exploring conical intersections, with SAECIS serving as a specialized extension for systems requiring the description of higher-lying states with significant double-excitation character.