Predictive Models for Removing Heavy Metal Water Pollutants with Biochar: Exploring Neural Networks and Machine Learning

This study evaluates the performance of hybrid Artificial Neural Network (ANN) models integrated with nature-inspired algorithms and machine learning (ML) techniques for predicting the efficacy of biochar in treating heavy metal pollutants in water. The Cuckoo Search Algorithm-optimized ANN (CSA-ANN) was the best-performing model on a dataset containing 353 data points and 17 variables. During training, it obtained a strikingly low Mean Squared Error (MSE) of 0.0053 and a high R 2 of 0.9976. Other prominent ANN configurations, including the Teaching-Learning-Based Optimization ANN (TLBO-ANN) and Particle Swarm Optimization ANN (PSO-ANN), also demonstrated predictive solid performance. The Gaussian Process Regression (GPR) model, using a Rational Quadratic kernel, was the best-performing ML model, with a validation MSE of 0.004 and an R² of 0.967. Ensemble methods, such as Custom Ensemble (LS Boost) and Bagged Trees, also showed competitive results, with MSE values of 0.002 and 0.015 and R² scores of 0.982 and 0.875, respectively. A feature importance analysis, using the Minimum Redundancy Maximum Relevance (MRMR) algorithm, identified C₀ (mmol/g), pH, and carbon percentage as the most critical factors, with importance scores of 1.4, 1.0, and 0.9, respectively. ANN models, particularly CSA-ANN, have limitations because of their complexity and need on large datasets, even if they offer superior accuracy in complex environmental forecasts. This work demonstrates that while ANN models are very successful, more straightforward machine learning models like GPR and robust ensemble approaches offer flexible, competitive, and interpretable alternatives for real-world applications.