Yapay Sinir Ağları Yaklaşımı İle Curuflarda Fosfor Kapasitelerinin İncelenmesi

Abstract

Yapay zeka, insanın düşünme yöntemini analiz ederek öğrenme, fikir üretme, iletişim kurma veya çıkarım yapma gibi yeteneklerini simüle etmeye çalışan uygulamalardır. Yapay sinir ağları (YSA), yapay zekanın önemli bir temsilcisi konumundadır ve sınıflandırma, öngörü ya da modelleme gibi birçok alanda başarı ile uygulanmaktadır. Deneysel çalışmaların sıklıkla kullanıldığı mühendislik uygulamalarında yapay sinir ağları yaklaşımının kullanımı giderek yaygınlaşmaktadır. Curufların kapasiteleri, metallerde istenmeyen empüritelerin üretim sırasında curuf fazına geçebilme yeteneklerinin bir ölçüsüdür. Empüritelerin çeşitlerine göre curuflarda farklı empürite kapasiteleri vardır. Fosfor kapasitesi de fosfor elementinin curuf fazına geçişinin bir ölçüsüdür. Fosfor, üretim aşamasında ortamdaki kısmi oksijen basıncına bağlı olarak fosfat ( PO_4^(3-)) ya da fosfür 〖(P〗^(3-)) iyonu halinde sistemde bulunur. Buna göre curuflarda fosfor kapasitesi, fosfat kapasitesi (C_(PO_4^(3-) )) ya da fosfür kapasitesi 〖(C〗_(P^(3-) )) olarak isimlendirilir. Curuflarda empürite kapasitelerinin hesaplanmasında kullanılan çeşitli matematiksel ve termodinamiksel modeller vardır. Birçok farklı alanda, farklı uygulamalar için kullanılan yapay zeka, curuflarda empürite kapasitesi hesaplamak için kullanılabilecek önemli bir alternatif olabilir. Özellikle son yıllarda gerçekleştirilen çalışmalarda yapay zekanın bir uygulayıcısı olarak kullanılan yapay sinir ağları da curuf kapasitelerinin tahmin edilmesinde önemli bir potansiyele sahiptir. Bu çalışmada, MATLAB R2011a yazılımı kullanılarak literatürde yer alan deneysel çalışmalardan alınan veriler ile yapay sinir ağı modelleri oluşturulmuştur. Yang ve diğerlerinin, 2011 yılında gerçekleştirdikleri 1973 K’de CaO – SiO2 – MgO – FeO – Fe2O3 – MnO – Al2O3 curuf sisteminin fosfat kapasitesi (C_(〖 PO〗_4^(3-) )) tahminine ait deneysel verileri ile YSA modeli oluşturulmuş, elde edilen çıktılar ile aynı çalışmada yer verilen termodinamiksel bir model olan “İyon ve Molekül Birlikte Varoluşu (IMCT)” teorisi çıktıları karşılaştırılmıştır. Maramba tarafından 2007 yılında gerçekleştirilen, 1773 K sıcaklıkta SiO2 – Al2O3 – FeO – CaO – MgO – MnO curuf sisteminde farklı kompozisyona sahip sentetik curufların CO atmosferi altındaki fosfür kapasiteleri 〖(C〗_(P^(3-) )). ölçümünün deneysel verileri ile YSA modeli oluşturulmuş, elde edilen çıktılar ile aynı çalışmada yer alan deneysel verilere göre İstatistiksel Analiz Yazılımı (SAS) yardımı ile fosfür kapasitesini hesaplamak için oluşturulan matematiksel bir regresyon modeli çıktıları karşılaştırılmıştır. Literatürde yer alan fosfor kapasitesi çalışmalarından elde edilen veriler ile farklı silikat ve halojenür sistemlerine ait fosfor kapasitesi tahminleri gerçekleştirilmiştir. CaO – CaF2 – Al2O3 üçlü sisteminin faz diyagramındaki 1773 K sıcaklıkta sıvı bölgesi üzerinde eşdeğer fosfor kapasitesi eğrileri çizimi gerçekleştirilmiştir.

Artificial intelligence is the intelligence of software or machines that simulates learning, generating ideas, communicating or syllogizing abilities of human brain. An Artificial Neural Network (ANN) is a significant sub-branch of the artificial intelligence that is mostly used for solving classification and prediction problems. Artificial Neural Network approach is becoming very common in experimental researches of many engineering disciplines. There are several applications with the ANN on various fields such as, function approximation, regression analysis, classification, pattern recognition, data processing, clustering, source separations, filtering, time series prediction, robotics and controls. With using these various problem solving fields, the ANN is commonly used in medical diagnosis, financial predictions, industrial controls, data mining, sales and marketing, operational analysis, education, science etc. Unlike digital computers, the ANNs have inducting ratiocination. Trained samples are presented to the model as inputs and outputs, but rules are generated by user of the ANN. If the information or data is noisy, partial or the rules are unkown, error tolerances can be carried out. However, the digital computers are much more faster than the ANNs. Data and algorithms in the digital computers are more accurate than the data and algorithms in the ANNs. Because, an ANN is generated by practises of the data taken from. Main advantage of the ANN is that they don’t need any mathematical models for processing. Also, the ANN do not need any rule bases. The ANN has “learning” ability and by using different learning algorithms, they are able to solve different kind of problems. The main drawback of the ANN is that they need a wide range of training for real world operations. Their powers of solving problems is limited with the data range that is able to use for input data. The other disadvantage of the ANN is that they are not able to do stability analysis. Except some network models, generally the ANN models cannot have advanced levels of adoptability. An artificial neural network simply learns “information” that are presented as input data, then classifies or predicts outcomes via nodes and layers. ANN generates suitable algorithms for predicting reliable outputs. Artificial neural networks imitates a human brain. A human brain consist of approximately 12 billion nerve cell called “neurons”. A human body impulses a reaction via interactions of neurons. Comparably, an artificial neural network works as a neural system of a human body. Mainly, an ANN consist of an input layer, an output layer and hidden layers between input and output layers. This kind of structure is the most common neural network model that is called as Multi Layer Perceptrons. The inputs are fed into the input layer as “information” and they are get multiplied with weights before they send to the hidden layers. In the hidden layers they get processed by suitable functions. Finally, outputs are generated in output layer. The slags are complex compounds that consist of various forms of oxides, silicates, aluminates or borates. The major duty of slags is absorbing metallic or non-metallic inclusions in liquid metal. Beside, slags forms an interface between liquid metal and outer environment for preventing heat loss of metal. The components and characteristics of slags determines final properties of metal. Thus, the properties of slags might be differ by production process of desired metal. However, there are certain properties of slags for every production process. Firstly, the slags have to have a lower specific densites than liquid metal. Thus, the slag rises to the surface and covers the liquid metal. Secondly, volume of the slag has to be lower than the liquid metal, in order to lower the heat loss and energy loss of metal production process. There are mainliy three theories that express structures of slags. “Ionic Theory” supports that the basic oxides dissocates completely into ions. Then, acidic compounds combine with these ions to form complex compounds such as silicates, borates, etc. “Molecular Theory” supports that the structures of slags consist of either simple forms of molecules or complex forms of molecules that are formed by combining simple forms of molecules. “Coexistence Theory” supports that the structures of slags consist of both molecules and ions. Slag capacities denote measure of ability of slags for removing undesirable impurities while production of metal. There are several impurity capacities based on the impurity that liquid metals have. Phosphorus capacity denotes the ability of slag for removing phosphorus impurities in liquid metal. Depending upon partial oxygen pressure and temperature phosphorus favors either phosphate ( PO_4^(3-)) or phosphide 〖(P〗^(3-)) ions. Thus, the phosphorus capacity is distinctively named as either phosphate capacity (C_(PO_4^(3-) )) or phosphide capacity 〖(C〗_(P^(3-) )). In steelmaking process, high values of the partial oxygen pressure is reqiured. Thus, phosphorus forms as phosphate ion. For this reason, phosphorus capacity of steelmaking slags is calculated as phosphate capacity. In some special metal production processes, dephosphorization process is done with particular procedures. For instance, in ferromanganese and ferrochromium metal production processes, lower values of the partial oxygen pressure is required than ordinary steelmaking process for dephosphorization step. Because, high values of the partial oxygen pressure causes loss of substantially manganese or chromium in liquid metal to slag phase. In order to prevent loss of metal in production process of ferromanganese and ferrochromium, dephosphorization step needs to be done at reducing atmosphere. In such cases, phosphorus capacity is calculated as phospide capacity. There are several mathematical and thermodynamically model for calculating different impurity capacities. The artificial intelligence may a strong alternative method for calculating slag capacities. Especially in recent years studies with artificial neural network approach show that the ANN model has a grand potential in slag impurity capacity calculations. In this study, obtained experimental results on phosphorus capacities from variant studies in literature are used for modeling artificial neural networks via MATLAB R2011a software. According to experimental data that were taken from the phosphate capacity (C_(〖 PO〗_4^(3-) )) study on CaO – SiO2 – MgO – FeO – Fe2O3 – MnO – Al2O3 slag system at 1973 K, an ANN model was generated and output data of the ANN model was compared with output data of the thermodynamically model that was developed in same study as “Ion and Molecular Coexistence Theory (IMCT)”. Similarly, according to experimental data that were taken from the phosphide capacity 〖(C〗_(P^(3-) )) study on SiO2 – Al2O3 – FeO – CaO – MgO – MnO slag system at 1773 K, an ANN model was generated and output data of the ANN model was compared with output data of the regression model that was generated with Statistical Analysis Software pursuant to experimental data. Moreover, the ANN model was applied for the estimation of phosphorus capacities in binary and multi-component systems at different temperatures. Firstly, experimental data are taken from experimental studies, then these data are categorized as “silicates” and “halides”. The ANN model’s results plotted against experimental study results for comparison. Besides, iso-phosphorus capacity contours on liquid regions of CaO – CaF2 – Al2O3 ternary melt phase diagram at 1773 K temperature were generated and plotted by using the ANN model’s phosphorus capacity results.