Low-Dimensional Projections for Visualizing Energy Landscapes of Atomic Systems
Submission Track: Short Paper
Submission Category: AI-Guided Design
Keywords: Visualization, Projection, Energy landscape, Optimizer, Low-dimensional, principal component analysis
TL;DR: Proposed a method to visualize atomic energy landscapes in low dimensions, validated by Hessian analysis. Aids in understanding landscape and optimizers' trajectory.
Abstract: Understanding the energy landscape is key to discovering materials with targeted properties since the landscape encapsulates system's complete thermodynamic and kinetic behavior,
including its non-equilibrium properties, such as relaxation and metastable phases. However,
the curse of dimensionality prohibits one from effectively visualizing the energy landscape. Here, we propose a method to
visualize the complex energy landscape of atomic systems. We demonstrate that the proposed low-dimensional
projection aligns well with the curvatures of the actual landscape, validated through Hessian
analysis. Further, we show that we can gain interesting insights into the behavior of different
gradient-based and machine-learned optimizers using the proposed visualization approach. We
hope that the study will enhance the understanding of the energy landscapes, and contribute to a fundamental understanding of the physics of materials and accelerate
their discovery.
Submission Number: 14
Loading