This document discusses methods for recovering metals from solid waste including pyrometallurgy, hydrometallurgy, and biological leaching. It also discusses how machine learning algorithms like random forests and artificial neural networks are used for solid waste management and resource utilization modeling. The document proposes using a gradient boosting decision tree and particle swarm optimization approach to rapidly evaluate metal recovery potential from coal fly ash without complex experiments. It describes constructing two models based on physicochemical and elemental properties to predict metal fractions and then calculates a metal recovery potential index.
This document discusses methods for recovering metals from solid waste including pyrometallurgy, hydrometallurgy, and biological leaching. It also discusses how machine learning algorithms like random forests and artificial neural networks are used for solid waste management and resource utilization modeling. The document proposes using a gradient boosting decision tree and particle swarm optimization approach to rapidly evaluate metal recovery potential from coal fly ash without complex experiments. It describes constructing two models based on physicochemical and elemental properties to predict metal fractions and then calculates a metal recovery potential index.
This document discusses methods for recovering metals from solid waste including pyrometallurgy, hydrometallurgy, and biological leaching. It also discusses how machine learning algorithms like random forests and artificial neural networks are used for solid waste management and resource utilization modeling. The document proposes using a gradient boosting decision tree and particle swarm optimization approach to rapidly evaluate metal recovery potential from coal fly ash without complex experiments. It describes constructing two models based on physicochemical and elemental properties to predict metal fractions and then calculates a metal recovery potential index.
Secondary metal recovery methods from solid waste include pyro
metallurgy (Long et al., 2010), hydrometallurgy (Kumari et al., 2016), and biological leaching (Yang et al., 2014). Metal recovery is widely carried out from primary minerals (Brar et al., 2022), but these primary resources are becoming progressively exhausted by long-term mining, and the cost and difficulty of mining are also increasing (Rai et al., 2021) Page 2 In recent years, machine learning (ML) has become increasingly popular due to its advanced data interpretation capabilities and high efficiency. Random forests, artificial neural networks, and other ML algorithms are widely used in solid waste generation (Kannangara et al., 2018), treatment (Guo et al., 2021), management (Lakhouit et al., 2023), resource utilization (Guo et al., 2021) modeling, and prediction (Cheah et al., 2022). Liu et al. (2022) investigated the environmental risks of heavy metal leaching from 160 incinerated fly ash samples based on an interpretable ML method. Page 3 Given these limitations, in this study, we proposed a novel strategy to rapidly and accurately evaluate the metal recovery potential of CFA without the need for complex experimental investigations. Dataset 1 contains a total of 500 data points from 10 CFA samples whose features include four types of physicochemical properties, elemental properties, total concentration, and fractions, with a total of 50 features. Specifically, the physicochemical properties comprised of 16 feature variables, such as the particle size D10 and D50, oxides SiO2 and CaO, and loss on ignition, while the elemental properties included 32 feature variables, such as the atomic volume (See Table S1 for de tails). he water-soluble frac tion (F1), acid-soluble fraction (F2), reducible fraction (F3), oxidizable fraction (F4), and residual fraction (F5) ( F1 represents metal elements that can easily be removed using ultrapure water (Zhang et al., 2017). F2 comprises metal elements bound to car bonates that can pass into the water column. F3 consists of metal ele ments bound to iron and manganese oxides, while F4 represents the metal fraction bound to organic matter and sulfide under oxidizing conditions (Morillo et al., 2004). F5 includes metal elements that are associated with crystalline minerals. Page 4 In PSO, each particle in the particle swarm represents a possible solution to a problem. (1) Initialize the particle swarm Set the particle swarm size to T (generally ~20–40), randomly initialize the velocity and position of each particle in the velocity in terval and search space, Gradient boosting decision tree The gradient boosting decision tree (GBDT) method is an ensemble learning boosting algorithm that only uses decision trees as weak learners and combines the addition model and forward-step algorithm for iterative optimization. This algorithm involves three main concepts: decision tree, gradient boosting, and shrinkage 2) Define the fitness function and solve individual extrema as the initial fitness value for each particle. (3) The initial fitness value is taken as the local optimal solution of each particle and saved as the optimal particle position. These optimal solutions are then compared to find the global optimal solution, whose position is recorded (Grobler and Engelbrecht, 2018). (4) Update the velocity and position of each particle according to Formulas 1 and 2 (Sousa-Ferreira and Sousa, 2016). vdd
i = ωvd i + c1r1( pd i − xd i ) + c2r2(pd g − xid)
Therefore, we used the gradient boosting regression algorithm from the ensemble module of the scikit-learn library to construct the prediction model (Pedregosa et al., 2011), which benefits from the ease of use and wide range of functionality of the Python programming language Particle swarm optimization Particle swarm optimization (PSO) is a probabilistic global optimi zation algorithm based on bird foraging behavior both datasets were initially divided into training and testing sets using a ratio of 9:1. The size of the testing set was then progressively increased from 10% to 40% at 5% intervals, and the model performance under different split ratios was evaluated using the default parameters in scikit-learn Page 5 The mobility and fractions of metal elements affect the metal re covery potential of CFA. Even if metals have the same total concentra tion but belong to different fractions, their recovery potential will differ to some extent. On this basis, this study proposed a new index to mea sure metal recovery potential (MRP), whose formula is as follows: MRP = ( ∑i=n1 Fi × Ei) × Tc(5) where n represents the number of metal element fractions from sequential extraction, Fi is the percentage of the ith fraction, Ei repre sents the metal recovery potential of the ith fraction, which can be determined by inspection or expert knowledge, and Tc is the total con centration of the specific metal element under investigation in CFA Performance evaluation Better model performance is indicated by R and R2 values closer to 1 and smaller MSE and MAE values. In addition, in terms of significance, P < 0.01 (0.05) was used as a threshold to indicate a strong correlation between the two variables, which can also imply high model prediction accuracy. Feasibility analysis of GBDT-PSO Repetitions and split ratio The process of determining the number of the repeated split of data sets is shown in Fig. 3a. When the number of repetitions was below 10, the model performance fluctuated significantly. . As the number of repetitions increased, the curve fluctuation decreased gradually, and the R2 values of datasets 1 and 2 varied slightly around 0.778 and 0.772, respectively. As shown, for dataset 1, the R2 values of the model for the training and testing sets were 0.759 and 0.795, respec tively, for a testing set size of 10% Page 6 The PSO algorithm was used for hyper-parameter tuning, with the optimal values shown in Table 1. The optimization model 1, con structed from dataset 1, required 2000 weak estimators; this was nearly eight times the number of estimators required for model 2, which was constructed from dataset 2. These large differences in the numbers of required weak evaluators indicate that the learning ability of model 1 was better Specifically, the R2 and R values of model 1 on the testing set after hyper-parameter tuning were 0.88 and 0.94, higher than the equivalent values of 0.79 and 0.90 achieved by model 1 before tuning and also higher than the R2 = 0.84 and R = 0.92 values of optimized model 2. In addition, the P values of the training set and testing set for the optimized model 1 were 3.92 × 10− 16 and 4.71 × 10− 17, respectively. Both values met the significance test level of P < 0.01, indicating a significant correlation between the predicted results and actual values elemental properties on metal fractions. However, many researchers have indirectly corroborated our findings by exploring the differences in the fractions of different metal elements. fractions were different for different metal elements. Using the BCR method, the proportion of Cd in water-soluble and exchangeable fractions was found to be highest with a value of 47.3%, while Cr and Co exhibited high residual fraction proportions with values of 42.8% and 40%, respectively Page 10 In the DAT sample, the concentrations of Al and Fe were also both very high, corresponding to values of around 15,096 mg/kg and 16,054 mg/kg, respectively. However, although the overall Fe concentration was higher, the proportion of Al reached 63.64% in fractions F1–F3, which exceeded that of Fe (20.17%); hence, the recovery potential of Al ranked second after that of Ca. he different chemical fractions determine the mobility of metal (Kim et al., 2002); in turn, this mobility affects the metal recovery po tential. These techniques can also be used to evaluate the recovery potential of different metals in solid waste based on accessible fractions (Van Herreweghe et al., 2003). However, existing experi mental methods are expensive, labor-intensive, and time-consuming, which hinders the study of effective metal extraction from solid waste. Thus, in this study, we proposed a novel metal recovery potential evaluation strategy that can yield economic and environmental benefits and is conducive to sustainable development. Specifically, two GBDT-PSO models were constructed to predict metal fractions (F1–F5) based on the sequential extraction method. The identified metal recovery potential of the DAT sample was the greatest, with an MRP of 43,311.70, whereas that of the JAP sample was the lowest, with an MRP of 10,568.25. An acid leaching test conducted by King et al. (2018) confirmed that higher Ca concentrations could improve the extraction efficiency of metal elements; thus, the combination of the above two factors could explain the greatest metal recovery potential of Ca to some extent.