Interaction from structure using machine learning: in and out of equilibrium.

Area of Science:

• Computational physics
• Statistical mechanics
• Machine learning applications

Background:

• Predicting interatomic potentials from structural data is a challenging inverse problem in classical interacting systems.
• Existing methods lack exact solutions for deriving potential information solely from structural configurations.

Purpose of the Study:

• To develop a machine learning (ML) model capable of accurately predicting pair potentials from given system configurations.
• To demonstrate the efficacy of artificial neural networks (NNs) in solving this inverse problem across various phases of matter.

Main Methods:

• Utilizing artificial neural networks (NNs) trained on structural data (pair correlation functions) to predict underlying pair potentials.
• Testing the NN model on diverse system configurations, including crystalline, liquid, and gas phases.
• Evaluating the model's performance on both equilibrium and non-equilibrium (active matter) systems.

Main Results:

• The ML model achieved highly accurate predictions of pair potentials, irrespective of the system's phase (crystalline, liquid, gas).
• The trained network effectively handled configurations from equilibrium simulations.
• The model successfully predicted effective interactions for active matter systems, aiding in understanding motility-induced phase separation (MIPS).

Conclusions:

• Machine learning provides a rapid and accurate solution for the inverse problem of predicting pair potentials from structural information.
• The developed NN approach is versatile, applicable to various physical systems and phenomena like MIPS.
• ML predictions offer insights into effective interactions in complex systems, including phase transitions in active matter.