Accessing the performance of CC2 for excited state dynamics: a benchmark study with pyrazine

This study benchmarks the performance of RI-CC2 for ultrafast excited state dynamics in pyrazine by implementing analytical gradients and nonadiabatic couplings in Q-Chem to drive both vibronic coupling models and neural network-accelerated on-the-fly simulations, successfully reproducing experimental population decay times and identifying key vibrational modes and dark state participation in the internal conversion process.

The Gist

Imagine a molecule of pyrazine (a ring-shaped chemical found in everything from coffee to DNA) as a tiny, energetic trampoline park. When a photon of light hits it, the molecule gets a sudden jolt of energy and starts jumping wildly. The big question scientists have been asking for decades is: How does this molecule calm down?

Specifically, how does it dump that extra energy so quickly (in less than a trillionth of a second) without breaking apart? This process is called "internal conversion."

This paper is like a high-tech detective story where the authors built a super-accurate simulation to watch this process happen in slow motion. Here is the breakdown of their adventure:

1. The Problem: The "Black Box" of Chemistry

For years, scientists have tried to simulate this using computer models. But the best models were either:

• Too slow: Like trying to calculate the weather for a whole planet using a 1990s calculator. They took too long to run.
• Too inaccurate: Like using a blurry map to navigate a maze. They missed crucial details.

The authors wanted to use a very precise mathematical tool called RI-CC2. Think of this tool as a "Gold Standard" microscope for molecules. It sees the electrons and atoms with incredible clarity. However, this microscope was too heavy to carry around for a long journey (a full simulation). It needed a way to move fast without losing its sharp vision.

2. The Solution: The "Smart GPS" (AI)

To solve the speed problem, the team did something clever. They didn't just run the slow, heavy microscope for the whole trip. Instead, they:

1. Took the heavy microscope and ran it for a few short, critical stops to gather high-quality data.
2. Trained an Artificial Neural Network (AI) on that data. Think of this AI as a Smart GPS that learned the terrain perfectly from the few stops the heavy microscope made.
3. Now, they could run the simulation using the Smart GPS. It moves at the speed of a sports car but still knows the exact path because it learned from the "Gold Standard" data.

3. The Journey: What They Discovered

Using this new "Gold Standard + Smart GPS" combo, they watched the pyrazine molecule relax. Here is what they found, using some fun analogies:

• The "Ghost" Room (The Dark State):
  For a long time, scientists thought the molecule jumped from one bright room (an excited state) to another bright room, skipping a "dark room" in between because it was invisible.
  The Discovery: The authors found that the molecule does visit the dark room (called the $A_{1u}$ state). It's not a ghost; it's a crucial stop on the highway. The molecule actually spends time there, bouncing back and forth before settling down.

• The Bouncers (Vibrational Modes):
  The molecule has different ways it can wiggle and vibrate (like a guitar string vibrating at different notes).
  The Discovery: The authors identified two specific "wiggles" (vibrational modes named $Q_{9a}$ and $Q_{8a}$) that act like bouncers at a club. These bouncers are the ones pushing the energy from one room to the next, controlling the rhythm of the molecule's relaxation. Interestingly, they found a new bouncer ($Q_{9a}$) that previous studies had missed.

• The Timing:
  The simulation predicted that the molecule dumps its energy in about 26 femtoseconds (that's 0.000000000000026 seconds).
  The Result: This matches almost perfectly with real-world experiments, which measured it at 22 ± 3 femtoseconds. It's like the simulation predicted the exact time a runner would cross the finish line, and the real runner did it in the same time.

4. Why This Matters

This paper is a "benchmark study," which is like a stress test for a new car engine.

• They proved that their new method (RI-CC2 + AI) works perfectly for small molecules like pyrazine.
• They created a massive, high-quality dataset (a "training manual") that other scientists can use to build even better AI models.
• They showed that with their new "Smart GPS" (AI), we can now simulate these complex dances for much larger molecules in the future, which could help us design better solar panels, medicines, or materials.

In a nutshell: The authors built a super-fast, super-accurate simulator to watch a molecule cool down. They found that a "hidden" state plays a major role in the process and identified the specific vibrations that drive the action. It's a victory for both physics and artificial intelligence.

Technical Summary

1. Problem Statement

The paper addresses the computational challenges in simulating ultrafast nonadiabatic dynamics (specifically internal conversion) in polyatomic molecules.

• The Bottleneck: Accurate modeling of these processes requires electronic structure methods that balance accuracy and cost. Second-order approximate coupled-cluster singles and doubles (CC2) is a promising candidate for this balance. However, its application to Trajectory Surface Hopping (TSH) dynamics has been hindered by the lack of analytical energy gradients and nonadiabatic coupling vectors (NACVs) in most software packages.
• The Benchmark System: Pyrazine is chosen as the benchmark system due to its rich history of experimental and theoretical study regarding ultrafast internal conversion ($S_2 \to S_1$). Key open questions include:
  1. Does the "dark" $A_{1u}$ ($n\pi^*$) state actively participate in the internal conversion process?
  2. Which specific vibrational modes drive the coherent population transfer and quantum beats observed in experiments?
  3. Can CC2 accurately reproduce experimental decay times (approx. 22 fs) despite previous numerical instabilities reported in similar systems (e.g., adenine)?

2. Methodology

The authors implemented a robust workflow within the Q-Chem software package to enable CC2-based dynamics:

• Electronic Structure (RI-CC2):

  • Implemented Resolution-of-the-Identity (RI) CC2 with analytical gradients and NACVs.
  • Addressed the non-Hermitian nature of CC2 by utilizing both left and right transition dipole moments to construct a symmetrized dipole matrix for diabatization.
  • Used the cc-pVDZ basis set.

• Dynamics Approaches:

  1. Reduced-Dimensionality Vibronic Coupling (VC) Model:
     • Constructed a Hamiltonian involving the three lowest excited states ($B_{2u}$, $A_{1u}$, $B_{3u}$).
     • Parameterized linear and quadratic coupling coefficients using RI-CC2 data.
     • Solved using exact quantum dynamics (Matrix Product State Quantum Dynamics, MPSQD) and Fewest-Switches Surface Hopping (FSSH).
  2. Full-Dimensional Ab Initio On-the-Fly Dynamics:
     • Performed direct TSH simulations using RI-CC2 gradients and NACVs.
     • Machine Learning Acceleration: To overcome the high cost of on-the-fly RI-CC2 calculations, a Diabatic Artificial Neural Network (DANN) was trained.
     • The DANN uses an equivariant graph neural network to predict diabatic Hamiltonians, forces, and NACVs.
     • Training utilized an active learning loop and Jacobian descent for multi-objective optimization (energies, gradients, and NACVs).

• Simulation Details:

  • Initial conditions sampled from a Wigner distribution at 300 K.
  • Simulations ran for 100–200 fs with 100 (ab initio) to 500 (DANN) trajectories.
  • Decoherence was handled via the Energy-Based Decoherence (EDC) correction.

3. Key Contributions

• Software Implementation: Successfully implemented analytical gradients and NACVs for RI-CC2 in Q-Chem, enabling its use in nonadiabatic dynamics for the first time in a robust manner.
• Novel Diabatization Scheme: Proposed an approximate diabatization method for non-Hermitian CC2 theories by symmetrizing the transition dipole moment matrix, allowing for accurate population analysis.
• ML-Driven Dynamics: Demonstrated a workflow where a DANN model, trained on RI-CC2 data, accelerates full-dimensional dynamics by orders of magnitude while maintaining high accuracy.
• High-Quality Dataset: Generated a comprehensive dataset of energies, forces, and NACVs for pyrazine, serving as a resource for future machine learning developments in excited-state dynamics.

4. Key Results

• Role of the Dark $A_{1u}$ State:
  • Both the VC model and full-dimensional dynamics confirm that the dark $A_{1u}$ state actively participates in the internal conversion.
  • Population transfer occurs rapidly ($<10$ fs) from the bright $B_{2u}$ state to both $A_{1u}$ and $B_{3u}$ states.
• Vibrational Mode Drivers:
  • Identified $Q_{9a}$ and $Q_{8a}$ as the key vibrational modes driving coherent population transfer between $A_{1u}$ and $B_{3u}$.
  • The $Q_{9a}$ mode is crucial for the initial $B_{2u} \to B_{3u}$ transfer and subsequent coherent dynamics.
  • The $Q_{8a}$ mode drives the oscillation between $A_{1u}$ and $B_{3u}$.
  • Note: The study challenges previous attributions of quantum beats solely to the $Q_1$ mode, showing that $Q_{9a}$ and $Q_{8a}$ exhibit coherent oscillations distinct from their ground-state frequencies.
• Decay Times and Spectra:
  • The on-the-fly dynamics reproduce the experimental $B_{2u}$ population decay time of 26 fs, which is in excellent agreement with the measured value of $22 \pm 3$ fs.
  • RI-CC2 overestimates vertical excitation energies by $\sim0.3$ eV (a known systematic error), but correctly reproduces spectral bandwidths and relative oscillator strengths when static disorder is included.
• Method Validation:
  • The DANN-accelerated dynamics showed excellent agreement with reference RI-CC2 calculations, validating the ML approach for reducing computational cost without sacrificing accuracy.

5. Significance

• Validation of CC2 for Dynamics: This study proves that RI-CC2 is a reliable and accurate method for simulating ultrafast nonadiabatic dynamics in polyatomic molecules, overcoming previous concerns about numerical instabilities near conical intersections.
• Resolution of Scientific Debate: The results provide strong theoretical evidence supporting the active role of the dark $A_{1u}$ state in pyrazine's photophysics, resolving ambiguities in previous experimental interpretations.
• Scalability: The successful integration of stochastic RI-CC2 (sRI-CC2) (mentioned in the conclusion) suggests that these high-accuracy dynamics can be extended to much larger molecular systems ($O(N^3)$ scaling), bridging the gap between high-level quantum chemistry and complex molecular dynamics.
• Resource Generation: The provided dataset and trained models offer a foundational benchmark for developing next-generation machine learning potentials for excited-state chemistry.