A neural network is a computational model inspired by the structure and function of biological neurons, consisting of interconnected layers of nodes that transform input data into output predictions. It is the foundational architecture of modern machine learning and deep learning.
A neural network is a computational model inspired by the way biological brains process information. The basic unit is the artificial neuron, which receives input values, multiplies each by a learned weight, sums the products, adds a bias, and passes the result through a nonlinear activation function (such as ReLU, sigmoid, or tanh). These neurons are organized into layers: an input layer that receives the data, one or more hidden layers that transform the data into increasingly abstract representations, and an output layer that produces the final prediction. The network's behavior is determined by its weights and biases, which are learned from data through a process called training. During training, the network's predictions are compared to the true values (for supervised learning), and the weights are adjusted using an optimization algorithm β typically gradient descent and backpropagation β to minimize the error.
Neural networks are classified by their architecture and training paradigm. Feedforward neural networks (FNNs), also called multi-layer perceptrons (MLPs), are the simplest form β information flows in one direction from input to output. Convolutional neural networks (CNNs) use convolutional layers that share weights across spatial locations, making them highly effective for image and signal data. Recurrent neural networks (RNNs) have connections that form cycles, allowing them to maintain a "memory" of past inputs β well-suited for sequential data, though modern practice increasingly favors LSTM, GRU, and transformer architectures. Autoencoders are trained to reconstruct their input, learning compressed representations useful for dimensionality reduction and anomaly detection. Generative adversarial networks (GANs) consist of two networks β a generator and a discriminator β trained in opposition, capable of generating realistic synthetic data. Physics-informed neural networks (PINNs) incorporate physical laws (such as differential equations) directly into the network architecture or loss function. Graph neural networks (GNNs) operate on graph-structured data and are used for systems with complex connectivity.
Neural networks are widely used in earthquake engineering for a range of applications. In structural health monitoring, neural networks are used for damage detection and localization from vibration data, for system identification, and for automated modal analysis. In seismic hazard analysis, they are used for ground motion prediction, for magnitude and distance estimation from early signals, and for classifying ground motion characteristics. In performance-based earthquake engineering, they serve as surrogate models that approximate the results of expensive nonlinear response history analysis with dramatically reduced computational cost. In computer vision for earthquake engineering, CNNs are used for crack detection, damage classification from imagery, and structural component recognition. Challenges include: (1) the need for large, high-quality datasets, which is a particular problem in earthquake engineering where damaging events are rare; (2) the "black box" nature of deep networks, which limits interpretability; (3) the risk of overfitting and the need for careful validation; (4) the difficulty of transferring models across structures, regions, and conditions; and (5) the computational cost of training large networks. Physics-informed approaches are an active area of research to address some of these challenges by incorporating known physical constraints into the model. In Iran, neural network research in earthquake engineering is growing rapidly, with applications in SHM, damage detection, ground motion prediction, and surrogate modeling, supported by the availability of data from Iran's seismic and strong-motion networks.