Abstract
Deep neural networks (DNNs) have achieved exceptional performance acrossvarious fields by learning complex nonlinear mappings from large-scaledatasets. However, they encounter challenges such as high computational costsand limited interpretability. To address these issues, hybrid approaches thatintegrate physics with AI are gaining interest. This paper introduces a novelphysics-based AI model called the "Nonlinear Schr\"odinger Network", whichtreats the Nonlinear Schr\"odinger Equation (NLSE) as a general-purposetrainable model for learning complex patterns including nonlinear mappings andmemory effects from data. Existing physics-informed machine learning methodsuse neural networks to approximate the solutions of partial differentialequations (PDEs). In contrast, our approach directly treats the PDE as atrainable model to obtain general nonlinear mappings that would otherwiserequire neural networks. As a physics-inspired approach, it offers a moreinterpretable and parameter-efficient alternative to traditional black-boxneural networks, achieving comparable or better accuracy in time seriesclassification tasks while significantly reducing the number of requiredparameters. Notably, the trained Nonlinear Schr\"odinger Network isinterpretable, with all parameters having physical meanings as properties of avirtual physical system that transforms the data to a more separable space.This interpretability allows for insight into the underlying dynamics of thedata transformation process. Applications to time series forecasting have alsobeen explored. While our current implementation utilizes the NLSE, the proposedmethod of using physics equations as trainable models to learn nonlinearmappings from data is not limited to the NLSE and may be extended to othermaster equations of physics.