Graphene, a two-dimensional (2D) material, holds significant potential for various engineering applications due to its unique properties. One promising application is energy-efficient water desalination, where nanoporous graphene membranes can significantly improve water flux in reverse osmosis (RO) processes compared to traditional polymeric membranes. The nanopore geometry is crucial for desalination performance; larger pores allow high water flux but poor ion rejection, while smaller pores exhibit the opposite. Finding the optimal nanopore geometry is computationally expensive, requiring extensive and time-consuming simulations. This research proposes an AI-driven framework to overcome this challenge. Deep reinforcement learning (DRL), which trains an agent to make optimal decisions, is combined with a convolutional neural network (CNN) to rapidly predict the performance of nanopores. The CNN is trained on a dataset generated from molecular dynamics (MD) simulations of various graphene nanopores, enabling the DRL agent to quickly evaluate its design choices. The aim is to design nanopores that maximize water flux while maintaining a high ion rejection rate, offering a significant advancement in water desalination technology.

The introduction extensively reviews existing literature on graphene's applications in various fields, highlighting its potential for gas separation, energy storage, and DNA sequencing. It also cites numerous studies on the use of nanoporous graphene for water desalination, emphasizing the challenges associated with finding optimal nanopore geometries due to the high computational cost of evaluating various designs using MD simulations. The review justifies the need for a faster, AI-based approach to accelerate the design process.

The research employs a two-pronged AI framework: a deep reinforcement learning (DRL) agent and a convolutional neural network (CNN)-based performance predictor. The DRL agent iteratively designs nanopores by removing carbon atoms from a graphene sheet, aiming to maximize a reward function that balances water flux and ion rejection rate. The CNN acts as a fast performance predictor, trained on a large dataset of graphene nanopores generated via MD simulations. Data augmentation techniques, such as flipping and translating existing pores, are used to significantly expand the training dataset for the CNN. The CNN, based on ResNet50 architecture, shows superior performance in predicting water flux and ion rejection compared to other models like XGBoost and VGG16. The DRL agent uses the CNN's predictions to guide its design decisions, enabling rapid exploration of the design space. The reward function uses a generalized logistic function to heavily penalize low ion rejection while rewarding high water flux. The DRL agent is trained using the deep Q-learning algorithm with experience replay, optimizing its policy to create nanopores with the desired properties. Molecular dynamics simulations are used to validate the performance of the AI-generated nanopores, comparing them to conventional circular nanopores. t-SNE is used to visualize the high-dimensional CNN-extracted features, demonstrating the model's ability to capture the relationship between nanopore geometry and performance. The MD simulations used the LAMMPS package with the SPC/E water model and Lennard-Jones potentials. Graphene nanopore area and perimeter are calculated using computer vision methods.

The AI framework successfully generated thousands of graphene nanopores within a week, a task that would take decades using traditional MD simulations alone. The DRL-created nanopores demonstrated significantly higher water flux compared to circular nanopores with similar ion rejection rates. Analysis revealed that the irregular shape with rough edges of the AI-designed pores is key to their superior performance. The t-SNE visualization confirmed that the CNN effectively learned to extract relevant features from the nanopore geometry to predict water flux and ion rejection. MD simulations validated the AI's predictions, showing that DRL-designed nanopores, despite having a larger area, outperformed circular pores in ion rejection while maintaining comparable water flux. The superior ion rejection was attributed to the creation of “ion-free zones” in the corners of the irregular nanopores, sterically hindering the passage of hydrated ions. The perimeter/area ratio was identified as a key shape parameter influencing desalination performance, with higher ratios in DRL-designed pores leading to better ion rejection.

The results demonstrate the effectiveness of combining DRL and CNN for accelerating the design and screening of nanoporous materials. The framework overcomes the limitations of traditional MD simulations by enabling rapid exploration of the design space, greatly reducing the time required to identify optimal nanopore geometries. The discovery of the importance of rough edges and high perimeter/area ratio offers valuable insights for the design of high-performance desalination membranes. The findings have implications beyond water desalination, suggesting that this AI framework could be adapted to design optimal materials for other applications, such as gas separation and energy storage.

This study successfully demonstrated the use of an AI framework combining DRL and CNN to efficiently discover optimal graphene nanopores for water desalination. The AI-designed nanopores significantly outperformed conventional designs, highlighting the potential of AI for accelerating materials discovery. Future research could focus on extending this framework to other 2D materials and exploring different optimization strategies within the DRL algorithm.

The study used a higher-than-normal seawater salinity in MD simulations for computational efficiency. This might limit the direct applicability of the findings to real-world seawater desalination. Further research is needed to validate the performance of the AI-designed nanopores under various conditions, including different pressures and salt concentrations. The study focused on a specific type of graphene; the applicability of the method to other 2D materials requires further investigation.