Abstract
We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neuralnetwork framework designed for scientific machine learning applicationsinvolving long temporal sequences. By reformulating the original MechanisticNeural Network (MNN) (Pervez et al., 2024), we reduce the computational timeand space complexities from cubic and quadratic with respect to the sequencelength, respectively, to linear. This significant improvement enables efficientmodeling of long-term dynamics without sacrificing accuracy orinterpretability. Extensive experiments demonstrate that S-MNN matches theoriginal MNN in precision while substantially reducing computational resources.Consequently, S-MNN can drop-in replace the original MNN in applications,providing a practical and efficient tool for integrating mechanisticbottlenecks into neural network models of complex dynamical systems. Sourcecode is available at https://github.com/IST-DASLab/ScalableMNN.