Materials modeling neural networks

A challenging task in material science as well as in pharmaceutical research is to custom tailor a compound s properties. George S. Hammond stated that the most fundamental and lasting objective of synthesis is not production of new compounds, but production of properties (Norris Award Lecture, 1968). The molecular structure of an organic or inorganic compound determines its properties. Nevertheless, methods for the direct prediction of a compound s properties based on its molecular structure are usually not available (Figure 8-1). Therefore, the establishment of Quantitative Structure-Property Relationships (QSPRs) and Quantitative Structure-Activity Relationships (QSARs) uses an indirect approach in order to tackle this problem. In the first step, numerical descriptors encoding information about the molecular structure are calculated for a set of compounds. Secondly, statistical and artificial neural network models are used to predict the property or activity of interest based on these descriptors or a suitable subset. 

We have already met one tool that can be used to investigate the links that exist among data items. When the features of a pattern, such as the infrared absorption spectrum of a sample, and information about the class to which it belongs, such as the presence in the molecule of a particular functional group, are known, feedforward neural networks can create a computational model that allows the class to be predicted from the spectrum. These networks might be effective tools to predict suitable protective glove material from a knowledge of molecular structure, but they cannot be used if the classes to which samples in the database are unknown because, in that case, a conventional neural network cannot be trained. 

A more common use of informatics for data analysis is the development of (quantitative) structure-property relationships (QSPR) for the prediction of materials properties and thus ultimately the design of polymers. Quantitative structure-property relationships are multivariate statistical correlations between the property of a polymer and a number of variables, which are either physical properties themselves or descriptors, which hold information about a polymer in a more abstract way. The simplest QSPR models are usually linear regression-type models but complex neural networks and numerous other machine-learning techniques have also been used. 

A neural-network-based simulator can overcome the above complications because the network does not rely on exact deterministic models (i.e., based on the physics and chemistry of the system) to describe a process. Rather, artificia] neural networks assimilate operating data from an industrial process and learn about the complex relationships existing within the process, even when the input-output information is noisy and imprecise. This ability makes the neural-network concept well suited for modeling complex refinery operations. For a detailed review and introductory material on artificial neural networks, we refer readers to Himmelblau (2008), Kay and Titterington (2000), Baughman and Liu (1995), and Bulsari (1995). We will consider in this section the modeling of the FCC process to illustrate the modeling of refinery operations via artificial neural networks. 

An artificial neural network (ANN) model was developed to predict the structure of the mesoporous materials based on the composition of their synthesis mixtures. The predictive ability of the networks was tested through comparison of the mesophase structures predicted by the model and those actually determined by XRD. Among the various ANN models available, three-layer feed-forward neural networks with one hidden layer are known to be universal approximators [11, 12]. The neural network retained in this work is described by the following set of equations that correlate the network output S (currently, the structure of the material) to the input variables U, which represent here the normalized composition of the synthesis mixture ... 

The set of equations formed by fheseequations is then solved numerically. Such models have been used extensively to describe breakthrough curves onto activated carbon of mono-component solutions of metal ions, micro-organic compounds or dyes [20-22], Some studies have demonstrated that they could be used to model binary namic adsorption [23] but they may not be applied in the case of complex multi-solute solutions. In addition, they do not take into account the pore characteristics of activated carbon materials, which are known to influence strongly the adsorption of micro-orgaiucs. In these cases, statistical tools like neural networks may be used in order to introduce such parameters as explicative variables. 

The adjusted weekly values can now be used for multi-parametric regression using neural networks. The objective is to define the non-linear relation for the amount of ZIC produced in the presses and centrifuges. It is also aimed to model the contents of Zn, Fe, Pb, SiO and AI2O3 in the ZIC as a function of the chemical composition (Zn, Fe, Pb, Si02 and AI2O3) of the feed material. 

MSE values for different network architectures are presented in Table 5.10. From this table, it is seen that for mild steel material, increase in the number of neurons in the hidden layer beyond 4 does not improve the performance and thus network with 3-4-1 architecture is selected as the best network based on minimum MSE. For mild steel material, the maximum absolute percentage error is obtained as 0.69%. Thus, it is clear that the experimental values and ANN predicted values are very close. The comparative study of experimental fractal dimension and ANN model predicted fractal dimension is shown in Figure 5.10. From this figure also, it is clear that the predicted and experimental fractal dimension are very close to each other. The performance of the neural network is measured by carrying out the regression analysis. The correlation coefficient (R) for mild steel is 0.997 (Figure 5.11) and it is a good indication of the... 

---