1
Introduction

Pingfan Hu and Qingsheng Wang

Artie McFerrin Department of Chemical Engineering, Texas A&M University, College Station, TX, USA

Machine learning (ML) is a method spanning a broad array of disciplines, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and others. Furthermore, it is the core subset of artificial intelligence (AI). The term “machine learning” was first proposed in 1959 by Arthur Samuel (Samuel 1959). Machine learning algorithms can build mathematical models based on training data to make predictions or decisions without being explicitly programmed to do so. Bayesian and Laplace's derivations of least squares and Markov chains, which date back to the seventeenth century, have previously constituted the tools and foundations widely used in ML (Andrieu et al. 2003). Since then, the ML algorithms have developed tremendously and have been widely applied in various aspects of scientific research and everyday life. These include data mining (Mitchell 1999), computer vision (Voulodimos et al. 2018), natural language processing (Cambria and White 2014), biometric recognition (Chaki et al. 2019), medical diagnosis (Bakator and Radosav 2008), detection of credit card fraud (Modi and Dayma 2017), stock market analysis (Chong et al. 2017), speech and handwriting recognition (Nassif et al. 2019), strategy games (Robertson and Watson 2015), and robotics (Pierson and Gashler 2017).

Deep learning (DL) is a relatively new branch within the field of ML. It is an algorithm that uses artificial neural networks (ANNs) as the architecture to characterize and learn data. The concept of DL originates from the research of ANNs, and a multilayer perceptron with multiple hidden layers is a DL structure (Lecun et al. 2015). DL forms a more abstract high‐level representation attribute category or feature by combining low‐level features to discover distributed feature representations of data. Several DL frameworks have been utilized, including deep neural network (DNN), convolutional neural network (CNN), and recurrent neural network (RNN).

The applications of ML algorithms in chemical safety and health studies date back to the mid‐1990s (Lee et al. 1995). Some research used basic ML algorithms in toxicity classification and prediction studies. For other fields such as hazardous property prediction and consequence analysis, the implementation of ML/DL algorithms did not emerge until the late 2000s (Pan et al. 2008; Pan et al. 2009). Chemical safety and health, although an important field, has rarely been investigated using interdisciplinary research with applied ML. This is because at the early development stage of ML/DL, the algorithm was relatively primitive, and its excellent predictive capabilities and accuracy were not widely verified and proven. Second, due to the lack of relatively simple and easy‐to‐use toolkits and the high skill requirements for algorithms and programming, the applications of ML/DL algorithms in chemical safety and health research have been limited. As a result, studies implementing ML have been relatively rare in the field of chemical safety and health in the late twentieth century and first decade of the twenty‐first century.

However, with the rapid advancement of AI and computer science in the past 10 years, the importance of ML/DL and their unparalleled advantages over traditional statistical methods and labor‐intensive work have drawn increasing attention and hence have developed significantly. There is also growing interest in expanding the application of ML/DL in the research field of chemical safety and health in academia.

In this book, ML fundamentals as well as popular ML/DL tools for the implementation of ML/DL in chemical safety and health research are introduced (Jiao et al. 2020a). For the applications of ML/DL, the book describes flammability characteristics predictions using quantitative structure–property relationship modeling (Chapter 3), consequence prediction using quantitative property–consequence relationship modeling (Chapter 4), ML involving process safety and asset integrity management (Chapter 5), and ML for process fault detection and diagnosis (Chapter 6). Furthermore, the book describes intelligent methods for chemical emission source identification (Chapter 7), ML and DL applications in medical image analysis (Chapter 8), predictive nanotoxicology: nanoinformatics approach for toxicity analysis of nanomaterials (Chapter 9), ML in environmental exposure assessment (Chapter 10), and air quality prediction using ML (Chapter 11). This book provides useful guidance for researchers and practitioners who are interested in implementing ML/DL related to chemical safety and health. This book is an excellent reference for readers to find more information about novel ML/DL tools and algorithms.

1.1 Background

Author Tom Mitchell provides a modern definition of ML as follows: “A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P improves with experience E” (Jordan and Mitchell 2015). In general, there are three types of ML: supervised learning, unsupervised learning, and reinforcement learning.

Supervised learning learns a function from a given training data set. When new data (validation/test data) comes, it can predict the results based on the function. The training set requirements for supervised learning include inputs (features) and outputs (targets). The targets in the training set are already labeled (with specific experimental/simulation values). Common supervised learning algorithms include regression and classification algorithms. While some algorithms are only capable of classification analysis (e.g. linear discrimination analysis, naive Bayes classification), most of them (e.g. k‐nearest neighbor, random forest) are able to conduct both classification analysis and regression analysis (James et al. 2017; Witten et al. 2017).

The difference between supervised learning and unsupervised learning is whether or not the target of the training set is labeled. Compared with supervised learning, the training set of unsupervised learning has no artificially labeled results. Common unsupervised learning algorithms can be used for clustering (James et al. 2017; Witten et al. 2017). There is also semi‐supervised learning, which combines elements of supervised learning and unsupervised learning. The algorithm for semi‐supervised learning gradually adjusts its behavior as the environment changes.

For DL, the original work on neural networks was published by Warren McCulloch and Walter Pitts in 1943 (McCulloch and Pitts 1943). They introduced the McCulloch–Pitts neural model, also known as the “linear threshold gate.” As the first computational model of a neuron, the McCulloch–Pitts neural model is very simplistic, generating only a binary output. The weights and threshold require hand‐tuning. In the 1950s, the perceptron became the first model with the capability to autonomously learn the optimal weight coefficients, allowing the training of a single neuron (Rosenblatt 1958). With the help of the backpropagation algorithm, neural networks began to be trained with one or two hidden layers (Rumelhart et al. 1986).

A single hidden layer neural network consists of three layers: input layer, hidden layer, and output layer. In the neural network that is trained with supervised learning, the training set contains values for the inputs x and target outputs y. The hidden layer refers to the fact that in a training set, the true values for these nodes are not observed. As shown in Figure 1.1, a notation for the values of the input features is a [0], where the term “a” stands for activation. It refers to the values that different layers of the neural network pass on to the subsequent layers. After the input layer passes on the values x to the hidden layer, the hidden layer in turn generates some sets of activations, a [1]. Finally, the output layer generates some value a [2], which is a real number that equals the value of ŷ. The hidden layer and output layer are associated with the parameters w and b. In order to compute the outputs (a) of the neural network, which is a sigmoid function of z (σ(z)), it is similar to operating repeated logistic regression. The calculations are shown in Eqs. 1.1 through 1.4. Besides the sigmoid function, other activation functions can be used to compute the hidden layer values. In modern neural networks, the default recommendation is to use hyperbolic tangent (tanh) or the rectified linear unit (ReLU).

Figure 1.1 Structure of a single hidden layer neural network.

In recent years, the ML community has determined that some cases can only be learned using DNNs rather than the single hidden layer neural networks (Hinton et al. 2006). DNNs with multiple hidden layers can use earlier layers to learn about low‐level simpler features and then use the later...