Artificial Intelligence-Based Design of Reinforced Concrete Structures Artificial Neural Networks for Engineering Applications

• Main Author: Hong, Won-Kee
• Format: eBook
• Language: English
• Published: Chantilly Elsevier Science & Technology 2023
• Edition: 1
• Subjects: Artificial intelligence-Engineering applications

Table of Contents:

• 6.4.3 Design verification -- 6.4.3.1 Forward designs -- 6.4.3.1.1 Training verifications -- 6.4.3.1.2 Design verifications -- 6.4.3.2 Reverse designs -- 6.4.3.2.1 Reverse Design Scenario 1 -- 6.4.3.2.2 Reverse Design Scenario 2 -- 6.4.3.2.3 Reverse Design Scenario 3 -- 6.4.3.2.4 Reverse Design Scenario 4 -- 6.4.3.2.5 Reverse Design Scenario 5 -- 6.5 Design of singly reinforced concrete beams (machine learning) -- 6.5.1 Feature selection-based machine learning for design of singly reinforced concrete beams -- 6.5.2 Interpretation of feature selections (for training structural data) -- 6.5.2.1 Overview of feature selection -- 6.5.2.2 Training results of machine learning based on forward design -- 6.5.2.2.1 Data for training -- 6.5.2.2.2 Training with nonchained method based on feature scores -- 6.5.2.2.3 Training with chained method based on feature scores -- 6.5.2.3 Design accuracies based on forward design -- 6.5.2.4 Reverse design implanting artificial neural genes on an input-side -- 6.5.2.4 Reverse design implanting artificial neural genes on an input-side -- 6.5.3 Rationale for feature selection -- 6.6 Recommendations and conclusions -- Reference -- 7 - Design of doubly reinforced concrete beams based on artificial neural network (deep learning) and regression models (ma ... -- 7.1 Introduction -- 7.2 Motivation of the artificial neural network-based design -- 7.2.1 Previous researches -- 7.2.2 Importance of the chapter -- 7.3 Deep neural networks for structural engineering -- 7.4 Generation of large structural datasets and network training -- 7.5 Design of doubly reinforced concrete beams based on artificial neural network -- 7.5.1 Design scenarios -- 7.5.2 Design of doubly reinforced ductile concrete beam -- 7.5.2.1 Forward design -- 7.5.2.2 Reverse design -- 7.5.2.3 Formulation of back-substitution method
• Front Cover -- Artificial Intelligence-Based Design of Reinforced Concrete Structures -- Artificial Intelligence-Based Design of Reinforced Concrete Structures: Artificial Neural Networks for Engineering Applicatio ... -- Copyright -- Contents -- Preface -- Acknowledgments -- 1 - Design of reinforced concrete beams and columns based on artificial neural networks -- 1.1 What can be learned from this book? -- 1.2 An evolution of artificial neural networks in civil engineering -- 1.3 Common machine learning versus artificial neural networks with deep learning using deep layers -- 1.4 Accuracy and interpretability of common artificial intelligence models -- References -- 2 - Understanding artificial neural networks: analogy to the biological neuron model -- 2.1 A learning and memory capability similar to that of the human brain -- 2.2 Activation functions -- 2.2.1 Why activation functions? -- 2.2.2 Activation functions for squashing the linear part of neurons -- 2.2.3 Types of activation functions -- 2.2.3.1 tanh(x) -- 2.2.3.2 Sigmoid -- 2.2.3.3 Rectified linear unit function -- References -- 3 - Factors influencing network trainings -- 3.1 Requirement for good training accuracies -- 3.1.1 Training with extrapolated datasets -- 3.1.2 A lack of feature indexes -- 3.1.3 Discontinuous output parameters -- 3.1.4 The following steps can also be taken to efficiently to avoid overfitting -- 3.1.5 Input conflict for reverse designs -- 3.2 Data initialization -- 3.2.1 Why initialization? -- 3.2.1.1 Vanishing and exploding gradient issues due to wide distribution of neural outputs -- 3.2.1.2 Weights narrowly distributed to prevent vanishing and exploding gradient issues -- 3.2.2 How to initialize neural network parameters effectively: avoiding vanishing gradients due to large standard deviations -- 3.2.3 Types of initializations
• 3.2.3.1 How initializations are performed -- 3.2.3.2 Initializations of Xavier (or Glorot) and He et al -- 3.3 Data normalization -- 3.3.1 Why normalization for network training? -- 3.3.2 How to normalize neural network parameters effectively -- 3.3.3 Verification of training -- 3.3.4 Recovery scale of original dataset -- 3.4 Multilayer perception -- 3.4.1 Understanding artificial neural networks with multiple layers and neurons -- 3.4.2 What are the neurons, weights, bias, and activation functions used in artificial neural networks for structural applications? -- 3.4.3 Feedforward networks connected by weights and biases -- 3.5 Training, validation, testing, and design -- 3.5.1 Conditions for good artificial neural networks -- 3.5.2 Validation of artificial neural network -- 3.6 Backpropagation for adjusting weights and bias -- 3.6.1 Why backpropagations? -- 3.6.2 Backpropagation minimizing cost functions -- 3.6.3 Chain rule for backpropagation -- 3.7 Conclusions -- References -- 4 - Forward and backpropagation for artificial neural networks -- 4.1 Gradient descent algorithm -- 4.1.1 Introduction -- 4.1.2 Problem example -- 4.1.3 Gradient descent for calculating a single fitting variable -- 4.1.3.1 Weight and bias -- 4.1.3.1.1 Step 1: Initialization -- 4.1.3.2 Loss function -- 4.1.3.2.1 Step 2: calculating MSE -- 4.1.3.3 Learning rate -- 4.1.4 Gradient descent to determine multiple fitting variables -- 4.1.4.1 Step 1: establishing an initial fitting line -- 4.1.4.2 Step 2: calculating a loss function as a function of two fitting variables -- 4.1.4.3 Step 3: calculating a gradient of a loss function with respect to weight and bias -- 4.1.4.4 Step 4: minimizing a loss function by converging gradient (a slope) descents to zero -- 4.1.4.5 Step 5: selecting a learning rate to calculate step size
• 5.6.3.3.3 Step 3: Training an ANN on ρt -- 5.6.3.3.4 Step 4: Train an ANN on ρc -- 5.7 Chained training scheme with revised sequence -- 5.7.1 Why chained training scheme with revised sequence? -- 5.7.2 CRS with revised training sequence to enhance training accuracies of PTM and CTS -- 5.7.3 Steps for CRS -- 5.7.3.1 Step 1: Training an ANN on CBM -- 5.7.3.2 Step 2: Training an ANN on ρt -- 5.7.3.3 Step 3: Training an ANN on b -- 5.7.3.4 Step 4: Training an ANN on ρc -- 5.7.4 CRS-trained accuracies based on features recommended by NCA only and extra features -- 5.7.5 Design verifications based on training methods -- 5.7.6 Verification charts -- 5.8 Conclusions -- Acknowledgment -- References -- 6 - Singly reinforced concrete beams based on regression models and artificial neural networks -- 6.1 Significance of the chapter -- 6.2 Generation of big data -- 6.2.1 Program to design a singly reinforced concrete section -- 6.2.2 Data generation code for a singly reinforced concrete section -- 6.3 Beam design by an ANN based on TED (training on entire inputs and outputs simultaneously) -- one forward problem and four rev ... -- 6.3.1 Design scenarios -- 6.3.2 Forward design -- 6.3.2.1 Forward training and design based on TED -- 6.3.2.2 Forward training and design based on TED and PTM -- 6.3.3 Reverse design -- 6.3.3.1 Design accuracies of reverse design based on TED -- 6.3.3.2 Reverse training and design based on TED, PTM, and CRS -- 6.3.4 Design accuracies with data qualities -- 6.4 Singly reinforced concrete beams based on shallow neural network -- 6.4.1 Reverse design scenarios (one forward problem and five reverse problems) -- 6.4.2 Formulation of training network for reverse designs -- 6.4.2.1 Direct method -- 6.4.2.2 Back-substitution method -- 6.4.2.2.1 Reverse Design 1 of Table 6.4.1a -- 6.4.2.2.2 Reverse Design 2 of Table 6.4.1b
• 4.1.4.6 Step 6: repeating steps 4 and 5 until a gradient descent shown in Eq. (4.1.4.3) becomes zero -- 4.1.5 Verification of predictions -- 4.2 A simple artificial neural network with forward propagation algorithm for a reinforced concrete beam -- 4.2.1 Forward propagation -- 4.2.2 Artificial neural networks based on backpropagation -- 4.2.2.1 Rate of changes in errors with respect to weight and bias -- 4.2.2.2 Calculation of mean squared error function -- 4.2.2.3 Updating weights connecting an output and neurons of the hidden layer -- 4.2.2.3.1 For weight ω21 -- 4.2.2.3.2 For weight ω22 -- 4.2.2.3.3 For weight ω23 -- 4.2.2.3.4 For weight ω24 -- 4.2.2.4 Updating weights connecting neurons of the hidden layer and inputs -- 4.2.2.4.1 For weight ω1 -- 4.2.2.4.2 Summary of weight update -- ω1 to ω5 -- 4.2.2.4.3 Summary of weight update -- ω6 to ω10 -- 4.2.2.4.4 Summary of weight update -- ω11 to ω15 -- 4.2.2.4.5 Summary of weight update -- ω16 to ω20 -- 4.2.3 Designing RC beams based on reverse artificial neural networks -- 4.3 Conclusions -- 5. Training methods: designs based on training entire data, parallel training method, chained training scheme, and chained tra ... -- 5.1 Past studies -- 5.2 Significance of the chapter -- 5.3 Machine learning models versus deep layers based on artificial neural networks for structural engineering applications -- 5.4 Artificial neural networks and big data generation -- 5.5 Feature selection scores -- 5.6 Training methods, TED, PTM, CTS, and CRS -- 5.6.1 Training entire data -- 5.6.2 Parallel training method -- 5.6.3 Chained training scheme -- 5.6.3.1 Why chained training scheme? -- 5.6.3.2 Selection of feature indexes based on the feature selection scores -- 5.6.3.3 Steps for chained training scheme -- 5.6.3.3.1 Step 1: Training an ANN on CBM -- 5.6.3.3.2 Step 2: Training an ANN on b
• 7.5.2.3.1 Application to a reverse design with nine inputs (φMn, μφ, Mu, MD, ML, L, b, fy, and f'c) and nine outputs (h, d, ρrt, ρrc, ...