Representations in neural network based empirical potentials.

Area of Science:

• Computational materials science
• Machine learning in physics
• Quantum mechanics

Background:

• Accurately simulating atomic interactions is crucial for predicting material properties.
• Quantum mechanical simulations are computationally expensive.
• Machine learning offers efficient approximations for complex functions, like density functional theory.

Purpose of the Study:

• To investigate the "black box" nature of machine learning models in scientific applications.
• To demonstrate that machine learning models learn physical concepts, not just approximations.
• To provide insights into how neural networks model atomic interactions.

Main Methods:

• Training neural networks to reproduce density functional theory calculations.
• Utilizing dimensionality reduction techniques.
• Analyzing the internal representations of silicon atoms within the neural network.

Main Results:

• Machine learning models learn underlying physical principles during training.
• Dimensionality reduction reveals how neural networks represent atomic interactions.
• Insights into the learning process of neural networks for materials modeling.

Conclusions:

• Machine learning models can serve as more than mere approximations in scientific inquiry.
• Understanding the internal workings of these models enhances their reliability and interpretability.
• This research opens avenues for more transparent and insightful AI applications in materials science.