Prediction of Thermogravimetric Data in Bromine Captured from Brominated Flame Retardants (BFRs) in e-Waste Treatment Using Machine Learning Approaches

Abstract

The principal objective in the treatment of e-waste is to capture the bromine released from the brominated flame retardants (BFRs) added to the polymeric constituents of printed circuits boards (PCBs) and to produce pure bromine-free hydrocarbons. Metal oxides such as calcium hydroxide (Ca­(OH)2) have been shown to exhibit high debromination capacity when added to BFRs in e-waste and capturing the released HBr. Tetrabromobisphenol A (TBBA) is the most commonly utilized model compound as a representative for BFRs. Our coauthors had previously studied the pyrolytic and oxidative decomposition of the TBBA:Ca­(OH)2 mixture at four different heating rates, 5, 10, 15, and 20 °C/min, using a thermogravimetric (TGA) analyzer and reported the mass loss data between room temperature and 800 °C. However, in the current work, we applied different machine learning (ML) and chemometric techniques involving regression models to predict the TGA data at different heating rates. The motivation of this work was to reproduce the TGA data with high accuracy in order to eliminate the physical need of the instrument itself, so that this could save significant experimental time involving sample preparation and subsequently minimizing human errors. The novelty of our work lies in the application of ML techniques to predict the TGA data from e-waste pyrolysis since this has not been conducted previously. The significance of our work lies in the fact that e-waste is ever increasing, and predicting the mass loss curves faster will enable better compositional analysis of the e-waste samples in the industry. Three ML models were employed in our work, namely Linear, random forest (RF), and support vector regression (SVR), out of which the RF method exhibited the highest coefficient of determination (R 2) of 0.999 and least error of prediction as estimated by the root mean squared error (RMSEP) at all 4 heating rates for both pyrolysis and oxidation conditions. An 80:20 split was used for calibration and validation data sets. Furthermore, for showing versatility and robustness of the best-predicting RF model, it was also trained using all the data points in the lower heating rates of 5 and 10 °C/min and predicted on all the data points for the higher heating rates of 15 and 20 °C/min to again obtain a high R 2 of 0.999. The excellent performance of the RF model showed that ML techniques can be used to eliminate the physical use of TGA equipment, thus saving experimental time and potential human errors, and can further be applied in other real-time e-waste recycling scenarios.

Prediction of Thermogravimetric Data for Asphaltenes Extracted from Deasphalted Oil Using Machine Learning Techniques

“…In another study, GBR was seen to predict the bio-oil yields from hydrothermal liquefaction (HTL) of algae, than RF regression, and this helped optimize the HTL condition for different types of algae from the industry and also implement the ML-based optimization in biofuel production . In fact, in a previous work of our coauthor, they explored the ability of RF, SVR, and linear regression techniques for reproducing the TGA data of tetrabromobisphenol A (TBBPA) combined with Ca­(OH) 2 at 4 different heating rates (opposed to a single heating rate in our present work) under pyrolysis and combustion conditions. Here, we found that the RF model showed the best performance with high prediction accuracies of 0.999, while the SVR also showed good performance with R 2 values of >0.9 for all conditions.…”

Prediction of Thermogravimetric Data in the Thermal Recycling of e-waste Using Machine Learning Techniques: A Data-driven Approach

“…Keeping in mind the discussion in the previous paragraph, we chose to apply an 80:20 split for the calibration/validation data sets with random sampling (data points chosen at random) for our ML models. This was also the industry standard and was chosen due to the large size of our data set . As can be seen in Figure , we applied an 80:20 split for calibrating and validating the models in scenario 1, where the ML models were trained by using the TGA data at each HR for both TBP and TBP + hematite samples.…”

“…Typical model compounds of BFRs include 2,4,6-tribromophenol (TBP) and tetrabromobisphenol-A (TBBA) since they contain a number of brominated phenolic species and aromatics . A plethora of works in literature have focused on the copyrolysis of TBBA with various metal oxides, − but very few have focused on TBP. , We would like to highlight a recent work by us, where thermogravimetric analysis (TGA) of TBBA combined with Ca­(OH) 2 was investigated and various machine learning (ML) techniques were applied to reproduce the TGA data . This work was an extension of a previous work also by Ali et al, where TGA data were obtained for samples of both pure TBBA and TBBA combined with Ca­(OH) 2 .…”

“…Predicting the TGA data for new input conditions such as HRs, time spent by the sample inside the chamber, heat supplied by the TGA instrument at different times, and temperatures of the sample and the chamber at different times is achieved by establishing linear and nonlinear relationships between these inputs and the output mass loss data. In our previous work, we compared the prediction ability of random forest (RF) and support vector (SVR) regression techniques as with ordinary least-squares (OLS)-based multiple linear regression (MLR) for reproducing the TGA data of TBBA copyrolyzed with Ca­(OH) 2 under both oxygen and nitrogen environments. These techniques are elaborated in the ML Methods Section for readers’ reference.…”

“…Furthermore, the mass loss curves for the reactant mixture were not reproduced in this work. Predicting TGA data poses significant challenges due to its multivariate nature, and except in our previous work on TBBA, this area has not been explored in the past for e-waste recycling. Ali et al obtained thermo-kinetic parameters from the TGA data of TBP combined with hematite using model-fitting and iso-conversional methods (model-free determination of kinetic parameters, that is, without the assumption of a reaction pathway), which are effective but time-consuming.…”