Guided Adaptive Diffusion: An Evolutionary Framework for Multimodal Atomistic Structure Prediction.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Atomistic structure prediction is crucial for understanding materials properties.
• High-dimensional potential energy surfaces pose significant challenges for traditional search algorithms.
• Existing methods often struggle with scalability as dimensionality increases.

Purpose of the Study:

• To develop a novel adaptive diffusion framework for efficient atomistic structure prediction.
• To enhance the search for global and local minima on complex potential energy surfaces.
• To overcome limitations of traditional algorithms in high-dimensional spaces.

Main Methods:

• Reinterpreting neural network-based denoising as an evolutionary search mechanism.
• Incorporating geometric constraints for physics-informed sampling.
• Employing a memetic approach combining diffusion models with gradient-based relaxation.

Main Results:

• Demonstrated ability to locate global and low-energy local minima for Lennard-Jones and gold clusters.
• Framework remains effective on high-dimensional potential energy surfaces.
• Maintained population diversity and search efficiency during optimization.

Conclusions:

• The adaptive diffusion framework offers a powerful new tool for atomistic structure prediction.
• The approach successfully handles complex, multimodal potential energy surfaces.
• This method represents a significant advancement in computational materials discovery.