Physics Informed Neural Networks (PINNs)

Physics Informed Neural Networks (PINNs) are a class of machine learning models that train neural networks based on physical constraints or laws. In order to detect patterns, conventional neural networks only use the input–output pairs that are supplied to them during training. They are determined by data. Conversely, PINNs combine data-driven learning with well-established physical principles in an effort to increase learning efficiency and generalisation. Because they can solve real-world problems with noisy data and often incomplete physics, physics-informed neural networks, or PINNs, are gaining popularity in many engineering domains. In PINNs, automatic differentiation is used to evaluate differential operators without discretization errors, and a multitask learning problem is defined to fit simultaneously observed data while adhering to the underlying laws of physics. This book has mostly focused on modifying the PINN through the use of different activation functions, neural network architectures, gradient optimisation techniques, and loss function structures. Numerous other applications have been demonstrated for PINNs, even though they have proven to be more useful in some scenarios when compared to more conventional numerical techniques such as the Finite Element Method (FEM). Still, there is potential for improvement, especially in terms of open theoretical problems.