Predicting adsorption ability of adsorbents at arbitrary sites for pollutants using deep transfer learning

Water pollution caused by heavy metal ions (HMIs) and organic pollutants poses significant threats to ecological systems and human health. Developing highly efficient adsorbents to remove these pollutants is crucial. The adsorption ability of an adsorbent depends on its active sites and their activity intensities, which are currently difficult to measure directly. Traditional methods like ab initio density functional theory (DFT) calculations can predict adsorption capacity, but they are computationally expensive and time-consuming, particularly when considering the vast configurational space of possible materials and active sites. Artificial intelligence (AI), particularly machine learning (ML), offers a promising alternative to accelerate this process. This study leverages the power of AI to predict the adsorption capacity of adsorbents at arbitrary sites with high efficiency. Machine learning methods have been used in material science to predict properties, but often require large datasets and/or extensive model training. To overcome these limitations, this research adopts transfer learning (TL), a powerful technique that transfers knowledge gained from one dataset to another related but different dataset. This allows the model to achieve high prediction accuracy with considerably less data and computational time. The study focuses on a representative two-dimensional (2D) material, graphitic-C3N4 (g-C3N4), and its adsorption characteristics toward three HMIs: Pb(II), Hg(II), and Cd(II). 2D materials, such as g-C3N4, offer several advantages, including enriched adsorption active sites due to defects and functional groups, along with a high surface area owing to their ultra-thin nature. The use of TL allows for the prediction of adsorption capabilities for Hg(II) and Cd(II) using a smaller dataset by transferring knowledge from a well-trained model for Pb(II). This approach significantly reduces computation time while maintaining high accuracy.

Numerous studies have explored the design and synthesis of various adsorbents for removing HMIs and organic pollutants from water. Existing methods for evaluating adsorption capacity include experimental measurements and computationally intensive ab initio calculations such as DFT. While these methods provide valuable insights, they have limitations. Experimental determination of adsorption at specific active sites is challenging. DFT calculations are accurate, but their computational cost restricts their applicability to a limited number of sites. In recent years, there has been increasing interest in applying AI and ML techniques to materials science and discovery to improve efficiency and reduce costs. Several studies have used machine learning to predict material properties, but challenges remain, particularly regarding data requirements and computational time. This study tackles these challenges by employing transfer learning, a technique that has shown promise in other fields but has seen limited applications in adsorption energy prediction.

This study utilizes a deep neural network (DNN) approach combined with transfer learning (TL) to predict the adsorption energies of three heavy metal ions (Pb(II), Hg(II), and Cd(II)) on the surface of graphitic-C3N4 (g-C3N4). The methodology involves the following steps: 1. **First-principles calculations (DFT):** Density functional theory (DFT) calculations using the Vienna ab initio Simulation Package (VASP) were performed to obtain a dataset of adsorption energies. For Pb(II)/g-C3N4, 7000 single-point adsorption energies were calculated at various sites on the g-C3N4 surface. For Hg(II)/g-C3N4 and Cd(II)/g-C3N4, 700 adsorption energies were calculated. The Perdew-Burke-Ernzerhof (PBE) exchange-correlation functional with the D3 dispersion correction was used to account for van der Waals interactions. 2. **Dataset creation:** The DFT calculations yielded datasets consisting of the adsorption energies (ΔE) and corresponding structural descriptors. The heavy metal ions were randomly placed on the g-C3N4 surface to ensure a representative sampling of adsorption sites. 3. **Structural descriptors:** A local environment matrix (LEM) was employed as the structural descriptor for the DNN model. LEM is suitable because it is an extensive, continuously differentiable approach and preserves all natural symmetries of the system. 4. **DNN model training:** The Deep Potential-Smooth Edition (DeepPot-SE) model was used, an end-to-end DNN-based potential energy surface (PES) model. This model was chosen due to its efficiency and accuracy in representing the potential energy surfaces of various systems. The model consists of three hidden layers, each with 20 nodes. The training was performed using the Adam optimizer with a batch size of 64. For the Pb(II)/g-C3N4 system, the model was trained from scratch using 6000 data points for training and 1000 for testing. For the Hg(II)/g-C3N4 and Cd(II)/g-C3N4 systems, transfer learning was used, initializing the model parameters with those from the trained Pb(II)/g-C3N4 model and fine-tuning with 600 training and 100 testing data points. 5. **Transfer learning:** Transfer learning was employed to leverage the knowledge learned from the Pb(II)/g-C3N4 model for predicting the adsorption energies of Hg(II) and Cd(II). This technique significantly reduces the amount of data needed to train the models for Hg(II) and Cd(II), while improving prediction accuracy. 6. **Experimental verification:** g-C3N4 was synthesized and its adsorption capacity towards the three HMIs was experimentally measured at two different initial concentrations (100 and 200 mg/L) using inductively coupled plasma (ICP) atomic emission spectrometry. The experimental results were then compared with the AI predictions to validate the model's accuracy.

The study's key findings demonstrate the effectiveness of the proposed AI-based approach for predicting adsorption capacity: 1. **High prediction accuracy:** The DNN model, trained with transfer learning, achieved root-mean-squared errors (RMSEs) less than 0.1 eV for all three HMIs (Pb(II), Hg(II), and Cd(II)) on g-C3N4. This accuracy is comparable to ab initio DFT calculations but with substantially reduced computational cost. 2. **Significant speedup:** The AI-based prediction is millions of times faster than DFT calculations, representing a considerable improvement in efficiency. 3. **Reduced data requirements:** Transfer learning significantly reduces the data needed for accurate predictions. Only 700 DFT-calculated adsorption energies were required for Hg(II) and Cd(II), compared to 7000 for Pb(II) when training from scratch. The transfer learning model achieved comparable or better accuracy to training from scratch with the smaller dataset, demonstrating its effectiveness. 4. **Prediction of adsorption site preferences:** The model accurately predicts the preferred adsorption sites. Analysis of the energy landscapes indicates that the HMIs are more likely to adsorb at the center of the g-C3N4 structure than at the edges. 5. **Consistent experimental validation:** Experimental measurements of the adsorption capacity of synthesized g-C3N4 for Pb(II), Hg(II), and Cd(II) showed adsorption capacity in the order Cd(II) > Hg(II) > Pb(II), which is consistent with the AI prediction. This confirms the validity of the proposed method. The mean adsorption energies for Pb(II), Hg(II), and Cd(II) were calculated as -1.664 eV, -1.695 eV, and -1.707 eV, respectively, indicating the relative adsorption abilities of Cd(II) > Hg(II) > Pb(II). The standard deviations highlight the range of adsorption energies at different sites on the surface.

The results demonstrate the feasibility and advantages of using deep neural networks combined with transfer learning for predicting adsorption abilities. The method significantly reduces computational costs and data requirements compared to traditional DFT calculations while maintaining high predictive accuracy. The successful prediction of adsorption site preferences and experimental validation further support the model's reliability. The ability to rapidly screen adsorbents for various pollutants using this approach could accelerate the design and development of new materials for environmental remediation. The observed order of adsorption (Cd(II) > Hg(II) > Pb(II)) is consistent with the relative strengths of their interactions with the g-C3N4 surface and can inform the design of more effective adsorbents. Future studies could explore the applicability of this approach to other types of adsorbents and pollutants, including organic contaminants. The methodology can be further improved by incorporating more sophisticated descriptors or expanding the training datasets.

This research successfully demonstrates an AI-driven approach for efficient and accurate prediction of the adsorption ability of adsorbents at arbitrary sites. Using deep neural networks and transfer learning, the model achieves accuracy comparable to computationally expensive DFT calculations but with a significant speedup and reduced data requirements. Experimental verification confirms the accuracy of the AI predictions. The findings highlight the potential of this method to accelerate materials discovery and optimization for environmental remediation, opening up possibilities for designing highly effective adsorbents for various pollutants. Future work should focus on extending this approach to a wider range of adsorbents and pollutants, and exploring the optimization of model parameters and structural descriptors for improved accuracy and efficiency.

The study focuses on a specific 2D material (g-C3N4) and three heavy metal ions. The generalizability of the model to other adsorbents and pollutants requires further investigation. The accuracy of the predictions depends on the quality and quantity of the DFT-calculated data used for training. The random sampling of adsorption sites may not capture all possible configurations, although efforts were made to ensure representative sampling. While the experimental validation supports the model's predictions, differences between experimental and predicted adsorption capacities may arise from factors not explicitly considered in the model, such as steric hindrance and interactions between HMIs.