New pollutant adsorption prediction method based on residual network model

Abstract

The invention discloses a new pollutant adsorption prediction method based on a residual network model. The method comprises the following steps: collecting and preprocessing data; a ResNet model is constructed; carrying out SHAP value analysis and model optimization; and predicting and verifying the adsorption of a new pollutant perfluorooctanoic acid. According to the ResNet model constructed by the method disclosed by the invention, R2 between a model predicted value and a true value of a new pollutant perfluorooctanoic acid is 0.990 and 0.993 respectively, mean square errors are 0.042 and 0.038 respectively, and the ResNet model has good generalization ability. According to the method, a large number of trial and error experiments are replaced with a high-precision prediction model, the experiment cost is reduced, the problem of model degradation occurring when a traditional neural network (ANN) is in a deepened level is solved, and the key problem that prediction of a traditional machine learning model fails in a small sample scene is broken through.

Description

[0001] This invention relates to the intersection of persistent organic pollutant control and artificial intelligence prediction technology, and specifically to a novel pollutant adsorption prediction method based on a residual network model. Background Technology

[0002] Perfluorooctanoic acid (PFOA), a typical persistent organic pollutant, is difficult to degrade due to its strong carbon-fluorine bonds. Concentrations exceeding 0.4 ng / L in surface water pose ecological risks and it has been included in my country's "List of Key Controlled New Pollutants." Metal-organic frameworks (MOFs) are ideal adsorbents for PFOA due to their ultra-high specific surface area and tunable pore size. Their metal clusters and functionalized groups (such as the amino group of UiO-66-NH2) can capture PFOA molecules through Lewis acid interactions and hydrogen bonds. However, the adsorption performance of MOFs is strongly coupled with variables such as specific surface area (SSA), median pore size (MPW), pH, and temperature. Traditional trial-and-error experiments are time-consuming and costly, relying on expensive detection equipment, resulting in low efficiency and difficulty in analyzing the interaction mechanisms of these variables. While some studies have employed artificial neural networks (ANNs) for prediction, these models are prone to degradation when the network depth is increased, limiting model depth and making it difficult to fully learn the complex mapping relationship between the deep structure of materials and adsorption performance, resulting in insufficient prediction robustness and generalization ability. At the same time, many machine learning models usually rely on large-scale, high-quality datasets to achieve optimal performance, but there are insufficient actual PFOA adsorption datasets (due to compound toxicity and synthesis costs), leading to prediction failure in small sample scenarios. More importantly, existing models cannot quantify the contribution of key variables (such as the synergistic effect of pH and MPW on adsorption capacity), which restricts the targeted optimization design of MOF adsorbents. Summary of the Invention

[0003] To address the technical problems mentioned above, this invention creatively introduces deep residual networks (ResNet) into the field of novel pollutant adsorption prediction for the first time, constructing a highly efficient and accurate method for predicting the adsorption performance of MOF materials for novel pollutants. This method achieves high-precision, multi-scenario prediction of the adsorption capacity of novel pollutants through structured feature table input.

[0004] To achieve the above objectives, the present invention provides the following technical solution: Step 1: Experimental Data Acquisition. MOF materials (zif-8, uio-67, uio-66-NH2) were synthesized and characterized, and their specific surface area (SSA), average pore size (APS), and other physical parameters were measured. Adsorption data were then obtained through batch adsorption experiments. Variables included: SSA, APS, median pore size (MPW), initial contaminant concentration (IC), reaction temperature (TEMP), pH, reaction time (T), and adsorption capacity (Q). e The collected data was then processed for outlier removal and normalization. Step 2, Deep Residual Network (ResNet) Model Construction: The ResNet architecture includes convolutional layers and pooling layers, residual blocks for in-depth feature learning, global average pooling layers, and fully connected layers; Step 3, Shapley addition method for interpreting SHAP value analysis and model optimization: core input variables are selected by calculating feature contribution based on game theory, the dataset is imported into ResNet models based on different optimization algorithms for training, and the optimal algorithm is selected based on the training results; Step 4, Prediction and Validation: The ResNet model predicts the adsorption of the new pollutant perfluorooctanoic acid by the MOFs material prepared in Step 1. The prediction results are compared with the actual adsorption experimental data to evaluate the accuracy and generalization ability of the model.

[0005] In step one, the normalization method is Min-Max normalization, which scales all input features to the [0,1] interval. The formula is as follows: ; in and These are the minimum and maximum values ​​of the data sample, respectively.

[0006] In step two, convolutional and pooling layers extract initial features, and multiple residual blocks perform deep feature learning. Each residual block contains skip connections and consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, satisfying a mapping relationship. y = F ( x )+ x Average pooling layer integration characteristics; The fully connected layer output dimension is 1, corresponding to an adsorption capacity Q. e The predicted value; when configuring the residual network structure, the residual block level can be customized.

[0007] Step three uses SHAP to quantify the contribution of each input feature to the prediction accuracy. After analysis and screening, the core input variables are reduced to five, including IC, pH, TEMP, T, and MPW. Optimization algorithms include Adaptive Gradient Descent (Adagrad), Adaptive Moment Estimation (Adam), Root Mean Square Propagation (RMSprop), and Stochastic Gradient Descent (SGD). The model selection criterion is the coefficient of determination R0. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ; The RMSE formula is: in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

[0008] The model performance evaluation metrics in step four include: coefficient of determination R² > 0.95 and root mean square error (RMSE) < 0.1.

[0009] 1. By replacing a large number of trial-and-error experiments with high-precision predictive models, the time and reagent costs required for MOF synthesis and screening, PFOA adsorption testing, and dependence on expensive detection equipment (such as high-precision chromatographs) are greatly reduced, thus accelerating the overall process of high-performance PFOA adsorbents from laboratory research and development to practical application.

[0010] 2. This invention proposes the introduction of deep residual networks into the field of novel pollutant adsorption prediction for the first time. The residual connection structure effectively solves the model degradation problem that occurs when traditional neural networks (ANNs) deepen, allowing for significant expansion of network depth. This enables the model to fully learn the highly complex nonlinear mapping relationship between PFOA adsorption capacity and the strongly coupled effects of multiple variables such as SSA, MPW, pH, and temperature, significantly improving prediction accuracy, model robustness, and generalization ability.

[0011] 3. The unique residual learning mechanism of deep residual networks, combined with structured feature table input, reduces the dependence on large-scale, high-quality training data. The model can effectively learn the inherent patterns of small sample data, and can still achieve efficient and reliable predictions even when PFOA adsorption experimental data is scarce, thus solving the key bottleneck of prediction failure of traditional machine learning models in small sample scenarios. Attached Figure Description

[0012] Figure 1 This is a flowchart of the ResNet-driven novel pollutant adsorption capacity prediction process of the present invention.

[0013] Figure 2 This is a topology diagram of the ResNet-18 model structure of the present invention.

[0014] Figure 3 Fitting curves for predicting the adsorption performance of PFOA in MOF materials (Adam+ResNet-18).

[0015] Figure 4 Fitting curves for predicting the adsorption performance of uranium (U(VI)) by biochar (Adam+ResNet-18).

[0016] Figure 5 Fitting curves for predicting CO2 adsorption performance of biochar (Adam+ResNet-18).

[0017] Figure 6 The results show the prediction of pollutant adsorption in MOFs under a small sample scenario (25% dataset + ResNet-18).

[0018] Figure 7 The impact of depth on prediction accuracy in traditional fully connected network (ANN) structures (1-18 layers).

[0019] Figure 8 To verify the prediction failure of ANN models in small sample scenarios (25% dataset + 18-layer ANN). Specific implementation methods

[0020] The embodiments of the present invention are described in detail below with reference to the accompanying drawings. These embodiments are exemplary and are used only to explain the present invention, and should not be construed as limiting the invention.

[0021] The ResNet, or Residual Neural Networks, mentioned in this invention, is a type of machine learning model.

[0022] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Example 1

[0023] Please refer to Figure 1 This invention provides a novel pollutant adsorption prediction method based on a residual network model, which includes the following steps: Step 1: Experimental Data Acquisition. MOF materials (zif-8, uio-67, uio-66-NH2) were synthesized and characterized, and their specific surface area (SSA), average pore size (APS), and other physical parameters were measured. Adsorption data were then obtained through batch adsorption experiments, with variables including SSA, APS, median pore size (MPW), initial contaminant concentration (IC), reaction temperature (TEMP), pH, and reaction time (T). Finally, UV-Vis spectroscopy was used to measure the residual contaminant concentration to determine the adsorption capacity (Q). e The collected data undergoes outlier removal and normalization. Min-Max normalization is used to scale all input features to the [0,1] interval, as shown in the formula: in and These are the minimum and maximum values ​​of the data sample, respectively. Step 2, ResNet Model Construction: The ResNet architecture starts with convolutional and pooling layers, followed by multiple residual blocks for in-depth feature learning, global average pooling layers, and fully connected layers. The convolutional and pooling layers extract initial features, while the multiple residual blocks perform in-depth feature learning. Each residual block contains skip connections and consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, satisfying a mapping relationship. y = F ( x )+ x Average pooling layer integration characteristics; The fully connected layer output dimension is 1, corresponding to an adsorption capacity Q. e The predicted value; the residual network structure is configured with 18 layers, and the structure diagram is as follows. Figure 2 As shown.

[0024] Step 3, SHAP value analysis and model optimization: Using game theory to calculate feature contribution, the core input variables were selected as IC, pH, TEMP, T, and MPW. The dataset was then imported into a ResNet model based on the Adam optimization algorithm for training. The evaluation metric was the coefficient of determination R. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ; The RMSE formula is: ; in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

[0025] Step 4, Prediction and Verification: 1. Input all the collected and processed data into the ResNet model for prediction. The results are as follows: Figure 3 As shown.

[0026] 2. Adsorption data of uranium (U(VI)) and CO2 by biochar were collected and processed from previous studies. A trained ResNet model was used to predict the adsorption of the collected and processed data. The prediction results are as follows: Figure 4 As shown in Figure 5, the R-squared value between the model prediction and the actual value is... 2 The values ​​are 0.990 and 0.993 respectively, with mean squared errors of 0.042 and 0.038. This indicates that the model has good generalization ability. Example 2

[0027] A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step two, the ResNet architecture begins with convolutional and pooling layers, followed by multiple residual blocks for in-depth feature learning, global average pooling layers, and fully connected layers. The convolutional and pooling layers extract initial features, while the multiple residual blocks perform in-depth feature learning. Each residual block contains skip connections and consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, satisfying a mapping relationship. y = F ( x )+ x Average pooling layer integration characteristics; The fully connected layer output dimension is 1, corresponding to an adsorption capacity Q. e The predicted value; the residual network structure is configured with 9 layers.

[0028] In step four, all the collected and processed data from the experiment are input into the ResNet model for prediction. Example 3

[0029] A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step two, the ResNet architecture begins with convolutional and pooling layers, followed by multiple residual blocks for in-depth feature learning, global average pooling layers, and fully connected layers. The convolutional and pooling layers extract initial features, while the multiple residual blocks perform in-depth feature learning. Each residual block contains skip connections and consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, satisfying a mapping relationship. y = F ( x )+ x Average pooling layer integration characteristics; The fully connected layer output dimension is 1, corresponding to an adsorption capacity Q.e The predicted value; the residual network structure is configured with 52 layers.

[0030] In step four, all the collected and processed data from the experiment are input into the ResNet model for prediction.

[0031] Please refer to Table 1. By comparing the prediction performance of 9-layer, 18-layer, and 52-layer residual networks, we found that the 52-layer model achieved the highest accuracy on the test set (R²=0.985, RMSE=0.042). This indicates that the ResNet model did not exhibit degradation, and increasing the network depth improves its ability to fit complex relationships in MOF adsorption. However, the performance difference between the 18-layer model (R²=0.983, RMSE=0.047) and the 52-layer model is less than 1%. It is worth noting that when the network depth increases to 52 layers, the number of model parameters increases to 1.9 times that of the 18-layer network, resulting in a training time increase of approximately 3.2 times, while GPU memory usage increases by 78%. Given that the 18-layer model already meets industrial-grade accuracy requirements (R²>0.95 and RMSE<0.1), and also has significant advantages in computational efficiency and hardware cost, 18 layers were selected as the optimal depth for the residual network.

[0032] Example 4

[0033] A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step three, the core input variables IC, pH, TEMP, T, and MPW were selected using game theory-based feature contribution calculation. The dataset was then imported into a ResNet model based on the Adagrad optimization algorithm for training, and the evaluation metric was the coefficient of determination R. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ; The RMSE formula is: ; in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

[0034] In step four, all the collected and processed data from the experiment are input into the ResNet model for prediction.

[0035] Example 5 A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step three, the core input variables IC, pH, TEMP, T, and MPW were selected using game theory-based feature contribution calculation. The dataset was then imported into a ResNet model based on the RMSprop optimization algorithm for training, and the evaluation metric was the coefficient of determination R. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ; The RMSE formula is: ; in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

[0036] In step four, all the collected and processed data from the experiment are input into the ResNet model for prediction.

[0037] Example 6 A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step three, the core input variables IC, pH, TEMP, T, and MPW were selected by calculating feature contribution based on game theory. The dataset was then imported into a ResNet model based on the SGD optimization algorithm for training, and the evaluation metric was the coefficient of determination R. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ; The RMSE formula is: ; in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

[0038] In step four, all the collected and processed data from the experiment are input into the ResNet model for prediction.

[0039] Please refer to Table 2. Systematic testing of residual network optimization algorithms shows that different optimization algorithms significantly affect model prediction performance. The Adam algorithm achieves the highest accuracy on the validation set (R²=0.983, RMSE=0.047), with a convergence speed 44% faster than Adagrad and better loss function stability than RMSprop. In contrast, the SGD algorithm fails to converge due to severe oscillations (R²<0.85). Further analysis reveals that the Adam algorithm effectively overcomes the gradient imbalance problem caused by differences in the scale of dependent variables in adsorption data by adaptively adjusting the learning rate, thus demonstrating a significant advantage in modeling complex nonlinear adsorption kinetics.

[0040]

[0041] Example 7 A novel pollutant adsorption prediction method based on a residual network model, with the same technical solution as in Example 1, differs in that: In step four, all the collected and processed data are randomly divided into four equal parts, and then input into the ResNet model for prediction.

[0042] The results are as follows Figure 6 As shown, even with scarce adsorption experimental data, the ResNet model can still achieve efficient and reliable predictions (Ri). 2 >0.85), which solves the key bottleneck of prediction failure of traditional machine learning models in small sample scenarios.

[0043] Comparative Example 1: In this comparative example, the technical solution is the same as in Embodiment 1, except that: In step two, the existing ANN model is used to predict the adsorption capacity Q. e Set the number of hidden layers to 1-9.

[0044] In step four, all the collected and processed data from the experiment are input into the ANN model for prediction.

[0045] Comparative Example 2: In this comparative example, the technical solution is the same as in Embodiment 1, except that: In step two, the existing ANN model is used to predict the adsorption capacity Q. e Set the number of hidden layers to 9-18.

[0046] In step four, all the collected and processed data from the experiment are input into the ANN model for prediction.

[0047] The results are as follows Figure 7As shown, as the number of layers in a traditional fully connected neural network (ANN) increases from 1 to 18, the model performance exhibits a non-monotonic change: peak accuracy is reached at the 9th layer (test set R²=0.950, RMSE=0.098), while the performance of the 18-layer model degrades significantly (R²=0.873, RMSE=0.184). This phenomenon of decreased accuracy due to increased depth is called neural network degradation.

[0048] Comparative Example 3: In this comparative example, the technical solution is the same as in Embodiment 1, except that: In step two, the existing ANN model is used to predict the adsorption capacity Q. e Set the number of hidden layers to 9.

[0049] In step four, all the collected and processed data are randomly divided into four equal parts, and then input into the ANN model for prediction.

[0050] The results are as follows Figure 8 As shown, when the amount of adsorption experimental data is reduced to 25%, the prediction performance of traditional fully connected neural networks (ANN) suffers a systematic collapse: the validation set R² of the ANN model drops to less than 0.6.

[0051] Therefore, this invention proposes a novel pollutant adsorption prediction method based on a residual network model. By introducing deep residual networks into the field of new pollutant adsorption prediction, the time cost required for MOF synthesis and screening and PFOA adsorption testing is greatly reduced. The residual connection structure effectively solves the model degradation problem that occurs when traditional neural networks (ANNs) are deepened. It also solves the key bottleneck of prediction failure of traditional machine learning models in small sample scenarios, and has broad application scenarios in the intersection of persistent organic pollutant control and artificial intelligence prediction technology.

[0052] It should be noted that the above embodiments are merely exemplary implementations of the present invention, and their detailed descriptions are intended to provide a clear understanding of the technical solutions, but should not be construed as limiting the scope of protection of the present invention in any way. For those skilled in the art, any substantially equivalent substitutions, adaptive modifications, or reasonable optimizations made without departing from the design concept and core technical principles of the present invention should be considered to fall within the scope of protection of the present invention.

Claims

1. A novel pollutant adsorption prediction method based on a residual network model, characterized in that, Includes the following steps: Step 1, Experimental Data Acquisition: Synthesize and characterize MOF materials (zif-8, uio-67, uio-66-NH2), and measure their specific surface area (SSA), average pore size (APS), and other physical parameters; Adsorption data were then obtained through batch adsorption experiments, with variables including: SSA, APS, median pore size (MPW), initial contaminant concentration (IC), reaction temperature (TEMP), pH, reaction time (T), and adsorption capacity (Q). e The collected data was then processed for outlier removal and normalization. Step 2, Deep Residual Network (ResNet) Model Construction: The ResNet architecture includes convolutional layers and pooling layers, residual blocks for in-depth feature learning, global average pooling layers, and fully connected layers; Step 3, Shapley addition method for interpreting SHAP value analysis and model optimization: core input variables are selected by calculating feature contribution based on game theory, the dataset is imported into ResNet models based on different optimization algorithms for training, and the optimal algorithm is selected based on the training results; Step 4, Prediction and Validation: The ResNet model predicts the adsorption of the new pollutant perfluorooctanoic acid by the MOFs material prepared in Step 1. The prediction results are compared with the actual adsorption experimental data to evaluate the accuracy and generalization ability of the model.

2. The novel pollutant adsorption prediction method based on a residual network model according to claim 1, characterized in that... The normalization method in step one is Min-Max normalization, which scales all input features to the [0,1] interval. The formula is as follows: in and These are the minimum and maximum values ​​of the data sample, respectively.

3. The novel pollutant adsorption prediction method based on a residual network model according to claim 1, characterized in that... In step two, convolutional and pooling layers extract initial features, and multiple residual blocks perform deep feature learning. Each residual block contains skip connections and consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, satisfying a mapping relationship. y = F ( x )+ x Average pooling layer integration characteristics; The fully connected layer has an output dimension of 1, corresponding to an adsorption capacity Q. e The predicted value; when configuring the residual network structure, the residual block level can be customized.

4. The novel pollutant adsorption prediction method based on a residual network model according to claim 1, characterized in that... In step three, SHAP is used to quantify the contribution of each input feature to the prediction accuracy. After analysis and screening, the core input variables are reduced to five, including IC, pH, TEMP, T, and MPW.

5. A novel pollutant adsorption prediction method based on a residual network model according to claim 1, characterized in that... The optimization algorithms in step three include Adaptive Gradient Descent (Adagrad), Adaptive Moment Estimation (Adam), Root Mean Square Propagation (RMSprop), and Stochastic Gradient Descent (SGD); the model selection criterion is the coefficient of determination R0. 2 and the root mean square error (RMSE) on the validation set, R 2 The formula is: ； The RMSE formula is: ； in , , These are represented as actual value, predicted value, and average value, respectively, and N represents the number of data points.

6. A novel pollutant adsorption prediction method based on a residual network model according to claim 1, characterized in that... The model performance evaluation metrics in step four include: coefficient of determination R² > 0.95 and root mean square error (RMSE) < 0.1.

Images