In a paper published in the journal Scientific Reports, researchers addressed rising urbanization and industrial emissions using advanced machine learning (ML) techniques to assess titanium dioxide (TiO2) photocatalytic degradation of air pollutants. Among the 13 models analyzed, lasso regression (LR2), decision tree (DT), and XG Boost (XGB) emerged with the highest R-squared (R2) values and showed remarkable accuracy, consistently displaying low mean absolute error (MAE) and root mean squared error (RMSE) values throughout the evaluation. Factors like dosage, humidity, and ultraviolet (UV) light intensity were key in predicting TiO2's effectiveness. This study highlights ML's potential in optimizing environmental remediation strategies.
Mechanism of photocatalytic degradation. Image Credit: https://www.nature.com/articles/s41598-024-64486-7
Related Work
Previous studies have discussed the problems caused by industrialization and urbanization, which raise air and water pollution. TiO2 is unique as it rapidly breaks down contaminants by photocatalysis. Traditional methods like UV-visible spectroscopy (UV–Vi) are slow and complex, while ML offers rapid and insightful predictions for optimizing TiO2's performance. ML, including artificial neural networks (ANNs), accelerates material property forecasts and catalyst design. Comparative analysis of ML models is essential for maximizing photocatalytic efficiency and understanding degradation mechanisms.
Methodology Overview
The methodology employed in this study began with sourcing a dataset of 200 experimental sets from previous research, focusing on TiO2 photocatalytic degradation. Seven key independent variables were identified: dosage, UV light intensity, humidity, and initial contaminant concentration, which influence the photocatalytic degradation rate (k).
As the logarithm of k shows generally small absolute values, it was calculated to enhance data analysis. The dataset underwent thorough statistical analysis, covering essential metrics such as mean, range (minimum and maximum), standard deviation, skewness, and kurtosis. These metrics collectively offered researchers a comprehensive insight into the data's distributional properties. These measures offered valuable insights into the dataset's distribution's central tendency, variability, and shape.
Data preprocessing played a crucial role in preparing the dataset for ML modeling, which involved cleaning the data, handling missing values, and standardizing numerical features using tools like standardscaler from scikit-learn. Furthermore, correlation studies were conducted to explore relationships among input variables, visualized through a correlation matrix. Strong correlations (> |0.5|) suggested potential multicollinearity concerns, but subsequent variance inflation factor (VIF) analysis affirmed negligible multicollinearity among variables.
Parametric studies further elucidated how dosage, humidity, UV intensity, and initial concentration affect photocatalytic degradation rates. This analysis highlighted varying impacts on k across different experimental conditions. Subsequently, ML model development commenced with a randomized test-train split of the dataset (7.5:2.5 ratio), ensuring robust training and validation phases. Each ML model's hyperparameters were optimized using grid search and tenfold cross-validation approaches to guarantee optimal performance.
The ML model's precision, robustness, and forecasting ability for photocatalytic degradation rates were assessed using several statistical measures. Overfitting assessment (OF), RMSE, MAE, correlation coefficients (R), R2, and Nash-Sutcliffe Efficiency (NSE) were among these metrics. Additionally, introducing the a20-index provided a novel engineering perspective on model reliability, highlighting predictions within a ±20% variation from experimental values.
Optimizing ML Models
The study concentrated on enhancing ML models' accuracy by optimizing hyperparameters for predicting TiO2 photocatalytic degradation rates. Superior prediction accuracy was shown by the XGB model, which performed exceedingly well, achieving high R2 values in training and testing. It highlighted its strong predictive capability and consistent performance in modeling the experimental data. Furthermore, R2 values between 0.923 and 0.927 demonstrated strong DT, LR2, and support vector regression (SVR) performance in the training and test phases. It underscores their capability to predict photocatalytic degradation rates effectively.
With minimal RMSE of 0.318 min-1/cm2 in the training phase and 0.450 min-1/cm2 in the test phase, statistical analysis showed that XGB maintained remarkable accuracy. Furthermore, XGB demonstrated a low MAE of 0.211 min-1/cm2 during training and 0.263 min-1/cm2 during testing, highlighting its reliability in forecasting experimental results. In contrast, linear regression models like ridge regression and LR1 consistently showed lower R2 values, indicating their limitations in accurately predicting the target variable.
The study utilized regression error characteristic (REC) curves to assess model performance visually. XGB displayed the smallest area over the curve (AOC) of 0.038 in training and 0.048 in testing. It confirmed XGB's superior predictive accuracy and ability to maintain consistency across different phases of model evaluation.
Score analysis ranked models based on overall performance, with XGB and DT emerging as top performers during training, scoring 103 and 90, respectively. While their scores slightly decreased in the test phase, XGB continued to lead with a total score of 98, highlighting its robust predictive power and reliability in real-world applications.
Conclusion
To sum up, this study assessed 13 ML models using a comprehensive dataset to predict the photocatalytic degradation rate of TiO2. The evaluated models encompassed linear methods, DT, random forests (RF), k-nearest neighbors (KNN), and boosting-based techniques. XGB stood out as the top performer, achieving the highest R2 values. Other models, such as DT, LR2, and SVR, also demonstrated strong predictive capabilities. In contrast, ANN, RR, and LR1 showed the lowest R2 values, indicating less effective performance in this context.
Regression and statistical analyses and regression error characteristics confirmed XGB and DT's superior predictive accuracy. The study highlighted the significant influence of dosage, UV light intensity, and humidity on the degradation rate. Overall, XGB, DT, and LR2 were identified as the most robust models, with recommendations for future work to include further training, testing with new inputs, and exploring additional experimental factors and new ML techniques.
Machine Learning Boosts Earthquake Prediction Accuracy in Los Angeles