Unlocking azobenzene isomerization mechanisms via an LLM agent-driven workflow integrating simulation, experiment,...

Area of Science:

• Photochemistry
• Molecular Switches
• Artificial Intelligence in Chemistry

Background:

• Bridged azobenzene derivatives are crucial photo-responsive molecular switches.
• Understanding their Z ↔ E isomerization mechanisms is difficult due to challenges in structure-spectrum relationships.
• Current spectroscopic and computational methods have limitations in providing comprehensive mechanistic insights.

Purpose of the Study:

• To develop and demonstrate an integrated, AI-driven workflow for studying the microscopic isomerization mechanisms of bridged azobenzenes.
• To establish clear structure-spectrum relationships for azobenzene photoisomerization.
• To provide a generalizable blueprint for AI-driven investigations of dynamic molecular systems.

Main Methods:

• Utilized a large-language-model (LLM) agent-driven workflow for literature-guided planning, ab initio molecular dynamics (AIMD) sampling, and density functional theory (DFT) spectral calculations.
• Employed robotic infrared/Raman measurements and interpretable machine learning, specifically an attention-based convolutional neural network (ATT-CNN).
• The ATT-CNN predicted the C-N[double bond, length as m-dash]N-C dihedral angle from vibrational spectra, achieving high accuracy (r = 0.99, MAE = 5°).

Main Results:

• The ATT-CNN accurately predicted the dihedral angle, demonstrating a strong correlation between vibrational spectra and molecular structure.
• Attention maps highlighted key spectral bands, enabling holistic interpretation of the isomerization mechanism.
• Transfer learning successfully extended the model's performance across different chemical environments and experimental datasets.
• LLM agents successfully planned and coordinated automated simulations and experiments, while human researchers focused on model development and interpretation.

Conclusions:

• The study presents the first LLM-agent-planned and orchestrated mechanistic investigation of azobenzene photoisomerization, integrating literature, theory, experiment, and machine learning.
• The developed workflow provides quantitative insights into azobenzene photoisomerization mechanisms.
• This AI-driven strategy offers a generalizable blueprint for future investigations of dynamic molecular systems.