Phase behavior of continuous-space systems: A supervised machine learning approach.

Area of Science:

• Computational physics and chemistry
• Soft matter physics
• Machine learning applications

Background:

• Predicting the phase behavior of complex fluids is computationally intensive using traditional molecular simulations.
• Supervised machine learning (ML) has shown promise for lattice models but requires adaptation for continuous-space systems.

Purpose of the Study:

• To extend supervised machine learning methods for identifying phase boundaries in continuous-space complex fluid systems.
• To develop and test a convolutional neural network (CNN) model for predicting phase diagrams of off-lattice models.

Main Methods:

• A novel convolutional neural network (CNN) model was developed using grid-interpolated molecular coordinates as input.
• The CNN model was trained and tested on two off-lattice models: the Widom-Rowlinson model and a freely jointed polymer blend.
• The method's ability to optimize phase transition searches using varying filter sizes was investigated.

Main Results:

• The ML approach accurately predicted phase diagrams for the tested off-lattice models, showing good agreement with established molecular simulation results.
• A significant advantage of the ML method is the elimination of critical slowing down, a common issue in traditional simulations near phase transitions.
• Incorporating intermediate structures near phase transitions into the training data was found to be crucial for accurate boundary prediction near critical points.

Conclusions:

• The proposed CNN-based ML method effectively determines phase boundaries for continuous-space complex fluid systems.
• This approach offers a computationally efficient and generalizable alternative to traditional molecular simulations for studying fluid phase behavior.
• The method's ease of implementation suggests broad applicability in complex fluid dynamics research.