Abstract
Data-driven discovery of "hidden physics" -- i.e., machine learning ofdifferential equation models underlying observed data -- has recently beenapproached by embedding the discovery problem into a Gaussian Processregression of spatial data, treating and discovering unknown equationparameters as hyperparameters of a modified "physics informed" Gaussian Processkernel. This kernel includes the parametrized differential operators applied toa prior covariance kernel. We extend this framework to linear space-fractionaldifferential equations. The methodology is compatible with a wide variety offractional operators in $\mathbb{R}^d$ and stationary covariance kernels,including the Matern class, and can optimize the Matern parameter duringtraining. We provide a user-friendly and feasible way to perform fractionalderivatives of kernels, via a unified set of d-dimensional Fourier integralformulas amenable to generalized Gauss-Laguerre quadrature. The implementation of fractional derivatives has several benefits. First, itallows for discovering fractional-order PDEs for systems characterized by heavytails or anomalous diffusion, bypassing the analytical difficulty of fractionalcalculus. Data sets exhibiting such features are of increasing prevalence inphysical and financial domains. Second, a single fractional-order archetypeallows for a derivative of arbitrary order to be learned, with the order itselfbeing a parameter in the regression. This is advantageous even when used fordiscovering integer-order equations; the user is not required to assume a"dictionary" of derivatives of various orders, and directly controls theparsimony of the models being discovered. We illustrate on several examples,including fractional-order interpolation of advection-diffusion and modelingrelative stock performance in the S&P 500 with alpha-stable motion via afractional diffusion equation.