MolOpt: Autonomous Molecular Geometry Optimization Using Multiagent Reinforcement Learning.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Machine Learning

Background:

• Traditional optimization problems often rely on manual algorithm selection and tuning, leading to inefficient trial-and-error processes.
• Learned optimizers, meta-learning, and learning to learn are emerging concepts addressing these inefficiencies.
• Molecular geometry optimization (MGO) typically involves hand-designed algorithms.

Purpose of the Study:

• To introduce MolOpt, a novel approach for autonomous molecular geometry optimization (MGO) using multiagent reinforcement learning (MARL).
• To develop a learned optimizer capable of performing MGO without reliance on pre-existing hand-designed optimizers.

Main Methods:

• Molecular geometry optimization (MGO) was framed as a multiagent reinforcement learning (MARL) problem.
• Each atom in the molecule was represented as an individual agent within the MARL framework.
• The MARL agents were trained to minimize forces acting on each atom to achieve MGO.

Main Results:

• MolOpt demonstrated generalization capabilities for MGO across various molecules (propane, pentane, heptane, hexane, octane) after training on simpler alkanes.
• Performance-wise, MolOpt surpassed the MDMin optimizer and matched the FIRE optimizer's efficiency.
• While effective, MolOpt's performance did not exceed that of the BFGS optimizer.

Conclusions:

• MolOpt serves as a proof-of-concept for the efficacy of MARL in autonomous molecular geometry optimization (MGO).
• The findings suggest that MARL offers a promising, novel approach for MGO, potentially opening new research avenues.
• This work highlights the potential for learned optimizers to revolutionize MGO processes.