Abstract The solution of fractional partial differential equations (PDEs) is an important topic in scientific computing. However, the traditional physics-informed neural networks (PINNs) have the problems of memory overflow and low computational efficiency when the derivative is discretized for a long time. Therefore, in this paper, we innovatively propose a framework of Laplace transform physics-informed neural networks (LT-PINNs), which is dedicated to solving the forward and inverse problems of Caputo-type fractional PDEs. The core of this method is to use the Laplace transform to construct the loss function, which skillfully avoids the dilemma that the fractional derivative operator in the traditional PINNs is difficult to operate effectively. Through the study of the benchmark problem of parameter a in a series of different scenarios, it is verified that LT-PINNs can predict the solution of Caputo-type fractional PDEs more accurately than fPINNs. The excellent performance of LT-PINNs in identifying inverse problems involving fractional order, convection and diffusion coefficient is further explored. At the same time, the effects of network structure, number of sampling points and noise on the LT-PINNs method are analyzed in detail. The results show that the method can predict the solution of the equation satisfactorily even under severe noise interference. The proposed LT-PINNs framework opens up a new path for efficiently solving fractional PDEs. It shows significant advantages in improving computational efficiency, reducing memory usage, and dealing with complex noise environments. It is expected to promote the further development of fractional PDEs in many fields.
Zhang et al. (Wed,) studied this question.