Accelerated Surface Hopping via Scaling the Spin--Orbit Coupling: Opportunities for Machine Learning

This paper investigates an accelerated surface hopping scheme for simulating ultrafast nonadiabatic processes by scaling spin-orbit couplings, demonstrating that while machine learning models can accurately predict potential energy surfaces and couplings to reduce computational costs, the final extrapolated time constants remain highly sensitive to fitting parameters, highlighting both the potential and current limitations of ML-enhanced reliability in this approach.

The Gist

Imagine you are trying to watch a movie of a tiny molecule changing its shape after being hit by a flash of light. This is a "non-adiabatic" process, where the molecule jumps between different energy states. The problem is that some of these jumps are incredibly slow—like watching a snail crawl across a continent. To see the whole movie, you need to simulate time scales that are currently impossible for standard computer models; they would take centuries to run.

To solve this, scientists use a "speed-up" trick. They artificially turn up the volume on the forces that cause the jump, making the snail run like a cheetah. They run the simulation at high speed, then mathematically slow the results back down to predict how long the real, slow process would take.

This paper is about testing that speed-up trick on a specific molecule called silaethylene (a cousin of ethylene, but with a silicon atom instead of carbon) and seeing if Artificial Intelligence (AI) can help make the results more reliable.

Here is a breakdown of what they did and found, using simple analogies:

1. The "Speed-Up" Problem

Think of the simulation like a race. To predict how long a marathon takes, you could run a sprint at 100x speed and then divide the time by 100. But to be sure your math is right, you need to run the sprint at different speeds (50x, 100x, 200x) and see if the pattern holds.

The authors found that to get a trustworthy answer, you need a huge number of "racers" (computer simulations called trajectories) for each speed. If you only have a few racers, the result is like guessing the winner of a race based on a coin flip—it's statistically shaky. Running enough racers is computationally expensive, like trying to hire a thousand runners just to time a single race.

2. The AI Solution (The "Cheat Code")

This is where Machine Learning (ML) comes in. Instead of calculating the complex physics for every single step of the race from scratch (which is slow), the team trained an AI to "memorize" the rules of the race.

• The Training: They showed the AI thousands of snapshots of the molecule moving.
• The Prediction: Once trained, the AI could predict the next move instantly, acting like a super-fast calculator.

The team used a clever technique called "Rotate-Predict-Rotate."

• Analogy: Imagine trying to teach a robot to recognize a cup. If you show it a cup upside down, it might get confused. So, before the robot looks at the cup, you rotate it to a standard position, let it make its guess, and then rotate the answer back to the original position. This helps the AI handle the 3D geometry of the molecule correctly.

3. What They Found

The team tested this AI on silaethylene, which has two main ways to relax:

1. The Fast Path: Dropping from a high-energy state to a lower one (Singlet to Singlet).
2. The Slow Path: A tricky jump to a "triplet" state (a different spin), which is very slow and hard to simulate.

The Good News:

• The AI was excellent at predicting the "Fast Path." The results matched the slow, super-accurate physics calculations almost perfectly.
• The AI successfully learned the "rules" of the molecule's energy landscape.

The Bad News (The Catch):

• When they tried to use the AI to predict the "Slow Path" (the triplet jump) and then use the speed-up math to guess the real time, things got messy.
• The Amplification Effect: The AI made tiny errors in its predictions. When they applied the "speed-up" math (scaling the forces), those tiny errors got blown up like a small crack in a dam turning into a flood.
• Because the math used to slow the results back down is very sensitive, the AI's tiny inaccuracies led to very different guesses for the final time constant. One method guessed the race took 468 seconds; the AI guessed 315 seconds.

4. The Conclusion

The paper concludes that while AI is a powerful tool that can run simulations much faster, you can't just blindly trust it for this specific "speed-up" method yet.

• The Recommendation: If you want to use AI here, don't try to run more speed-up scenarios with it. Instead, use the AI to run more racers within the same speed-up scenarios to get better statistics.
• The Warning: You must be very careful with how you train the AI. If the training data isn't perfect, the "speed-up" math will magnify those mistakes, giving you a confident but wrong answer.

In short: AI is a great engine for speed, but if the fuel (training data) has a tiny impurity, the "speed-up" math will make the car crash. The authors suggest a hybrid approach: use the slow, perfect physics for the most extreme speed-ups, and use the fast AI for the rest, but keep a very close eye on the results.

Technical Summary

1. Problem Statement

Nonadiabatic molecular dynamics (NAMD) simulations, particularly Trajectory Surface Hopping (TSH), are essential for modeling ultrafast photochemical processes like internal conversion (IC) and intersystem crossing (ISC). However, these methods are computationally expensive, limiting simulations to femtosecond timescales. Many processes of interest, such as ISC to triplet states in small molecules, occur on much longer timescales (picoseconds to nanoseconds), making them inaccessible to standard high-accuracy quantum chemical methods (e.g., MR-CISD).

To address this, Accelerated Surface Hopping (ASH) schemes scale the driving couplings (Spin-Orbit Couplings, SOCs, for ISC) by a factor $\alpha > 1$ to speed up the dynamics. The true time constant is then obtained by extrapolating results from multiple ensembles (with different $\alpha$) back to $\alpha = 1$.

• The Challenge: Reliable extrapolation requires running multiple large ensembles of trajectories. This creates a massive computational bottleneck. Standard practice often sacrifices the number of trajectories per ensemble ($n$) to afford multiple scaling factors, leading to poor statistical confidence and unreliable extrapolated time constants.
• The Opportunity: Can Machine Learning (ML) models, which are significantly faster than ab initio methods, be used to generate the necessary large trajectory ensembles to improve the statistical reliability of ASH extrapolations?

2. Methodology

Case Study: Silaethylene ($CH_2SiH_2$)

The authors selected silaethylene as a test case. It features a slow ISC process from the first excited singlet state ($S_1$) to the triplet state ($T_1$), driven by weak SOCs. The system involves two singlet states ($S_0, S_1$) and one triplet state ($T_1$).

Accelerated Surface Hopping (ASH) Protocol

• Scaling: SOCs were scaled by factors $\alpha = 15, 25, 35$.
• Extrapolation: Two fitting methods were tested to extrapolate the time constant $\tau$ to $\alpha=1$:
  1. Fit1: Assumes a fixed theoretical scaling $\tau^\alpha = A/\alpha^2$ (valid for two-state systems).
  2. Fit2: Fits data on a log-log scale ($\log \tau = C \log \alpha + D$), allowing the exponent to vary (expected to be near -2).
• Reference Data: High-level MR-CISD/SA-CASSCF(2,2) calculations were used as the ground truth.

Machine Learning Integration

The authors developed ML models to replace the expensive electronic structure calculations during dynamics propagation:

• Potential Energy Surfaces (PESs): Trained using Multi-State ANI (MS-ANI) models for $S_0, S_1$, and a separate ANI model for $T_1$.
• Non-Adiabatic Couplings (NACs): Trained using Kernel Ridge Regression (KRR) with a Rotate-Predict-Rotate (RPR) approach to handle vectorial properties and phase issues.
• Spin-Orbit Couplings (SOCs): Also trained using the RPR approach. Unlike NACs, SOCs are smooth, allowing the use of Cartesian coordinates (rotated via the inertia tensor) as descriptors without energy-gap multiplication.
• Training Strategy: Models were trained on a pool of 6000 points from reference trajectories (scaling factor 20), split into 5000 training and 1000 testing.

Simulation Setup

• Software: Newton-X (dynamics) interfaced with MLatom (ML predictions).
• Protocol: 200 trajectories were propagated for each scaling factor (15, 25, 35) for both the reference MR-CISD and the ML-driven simulations.
• Analysis: Populations were fitted to $A(1 - e^{-t/\tau^\alpha})$ using a multi-curve fitting strategy to ensure stability.

3. Key Contributions

1. Extension of RPR to SOCs: The study successfully adapted the Rotate-Predict-Rotate (RPR) scheme, previously used for NACs and dipoles, to learn Spin-Orbit Couplings directly from Cartesian coordinates, demonstrating its versatility for vectorial properties in NAMD.
2. Statistical Analysis of ASH: The authors rigorously quantified the uncertainty in ASH extrapolation. They demonstrated that at least 100 trajectories per ensemble are required to obtain a converged and reliable time constant, highlighting the computational cost of statistical accuracy.
3. Evaluation of ML in ASH: The paper provides a critical assessment of using ML to accelerate ASH. It identifies that while ML can reproduce populations within confidence intervals, the extrapolation process is highly sensitive to small errors in fitted time constants, particularly at high scaling factors.
4. Hybrid Workflow Proposal: The authors propose a hybrid strategy where high-scaling-factor trajectories (which are most sensitive to errors) are run with high-level theory, while lower-scaling-factor trajectories are generated using ML to expand the ensemble size efficiently.

4. Key Results

• ML Model Performance:

  • PESs: Achieved high accuracy ($R^2 > 0.95$) and correctly reproduced the $S_1 \to S_0$ deactivation pathways (Type T and Type B).
  • NACs & SOCs: The RPR models achieved satisfactory static accuracy ($R^2 > 0.9$ for NACs, $>0.98$ for $S_0-T_1$ SOCs).
  • Dynamics: ML-driven trajectories reproduced the $S_1$ decay and $T_1$ population growth within the 95% confidence interval of the MR-CISD reference.

• Extrapolation Discrepancies:

  • Despite good population agreement, the extrapolated time constants showed significant divergence between ML and MR-CISD results.
  • Fit1 Results: MR-CISD $\to$ 468 ps; ML $\to$ 315 ps.
  • Fit2 Results: MR-CISD $\to$ 225 ps; ML $\to$ 217 ps (closer, but likely coincidental due to slope compensation).
  • Root Cause: The discrepancy was traced to the ensemble with the highest scaling factor ($\alpha=35$). ML errors in SOC prediction were amplified at high $\alpha$, leading to noisier population transfers. When the $\alpha=35$ data was replaced with reference data, the ML extrapolation aligned perfectly with the reference.

• Scaling Factor Sensitivity: Increasing the number of scaling factors ($N$) did not improve results as much as increasing the number of trajectories per ensemble ($n$). Using very large $\alpha$ with ML models is risky due to error amplification.

5. Significance and Outlook

• Reliability of Accelerated Schemes: The study establishes that statistical convergence (high $n$) is more critical than the number of scaling factors for reliable ASH extrapolation.
• ML Limitations in ASH: While ML accelerates dynamics, its application in ASH requires extreme caution. The non-linear nature of the extrapolation (scaling $\alpha^{-2}$) means that small systematic errors in ML-predicted couplings at high scaling factors can lead to large errors in the final predicted lifetime.
• Future Directions: The authors advocate for Active Learning (AL) to iteratively refine ML models specifically in the regions of configuration space relevant to high-scaling dynamics. They suggest a hybrid workflow: use high-level theory for the most sensitive (high $\alpha$) data points and ML for the bulk of the lower $\alpha$ data to maximize efficiency without sacrificing extrapolation accuracy.

In summary, this work demonstrates that ML can successfully drive NAMD for silaethylene but highlights that direct substitution of high-level theory with ML in accelerated schemes is not trivial. The sensitivity of the extrapolation step necessitates rigorous validation and potentially hybrid approaches to ensure the physical reliability of the predicted time constants.