Abstract
Heterogeneous, multi-component materials such as industrial tailings or by-products, along with naturally occurring materials, such as ores, have been intensively investigated as candidate oxygen carriers for chemical-looping processes. However, these materials have highly variable compositions, and this strongly influences their chemical-looping performance. Here, using machine learning techniques, we estimate the performance of heterogeneous, multi-component materials as oxygen carriers for chemical-looping. Experimental data for 19 manganese ores chosen as potential chemical-looping oxygen carriers were used to create a so-called training database. This database has been used to train several supervised artificial neural network models (ANN), which were used to predict the reactivity of the oxygen carriers with different fuels and the oxygen transfer capacity with only the knowledge of reactor bed temperature, elemental composition, and mechanical properties of the manganese ores. This novel approach explores ways of dealing with the training dataset, learning algorithms and topology of ANN models to achieve enhanced prediction precision. Stacked neural networks with a bootstrap resampling technique have been applied to achieve high precision and robustness on new input data, and the confidence intervals were used to assess the precision of these predictions. The current results indicate that the best trained ANNs can produce highly accurate predictions for both the training database and the unseen data with the high coefficient of determination (R2 = 0.94) and low mean absolute error (MAE = 0.057). We envision that the application of these ANNs and other machine learning algorithms will accelerate the development of oxygen carrying materials for a range of chemical-looping applications and offer a rapid screening tool for new potential oxygen carriers.
Highlights
Chemical-looping combustion (CLC) and chemical-looping with oxygen uncoupling (CLOU) are state-of-the-art methods for heat and power production, which can achieve CO2 separation at low cost and with very low energy penalties [1]
Once optimised artificial neural network models (ANN) models were obtained with different hidden layers and number of neurons, the results showed that the data predicted by the ANN models fitted the experimental results well
We demonstrated an approach to predict the reactivity of manganese ores as oxygen-carrying materials in chemical-looping processes using machine learning (ML) algorithms named artificial neural networks (ANNs)
Summary
Chemical-looping combustion (CLC) and chemical-looping with oxygen uncoupling (CLOU) are state-of-the-art methods for heat and power production, which can achieve CO2 separation at low cost and with very low energy penalties [1]. In CLC, the fuel, here idealised as a pure hydrocarbon (CnH2m), is oxidised by the oxygen carrier, here denoted as a metal oxide (MyOx) to produce CO2 and steam [1,10] according to the following global reaction scheme:. Artificial neural networks are composed of input, hidden and output layers as well as a number of parallel-interconnected neurons in each layer. The input data are received by the neurons of the input layer and the output of each input neuron feeds into the neurons of a hidden layer. Yn are the output signals of the current layer and input signals of neurons of the layer; xn are the input signals of the network; win and bin are the values of weights and biases between the neurons, respectively. Yn are the output signals of the current layer and input signals of neurons of the layer; xn are the input signals of the network; win and bin are the values of weights and biases between the neurons, respectively. f is the activation function, where a sigmoid function is usually applied, as shown in Eq (8). f(x ) 1 1+e x (8)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
