Gaussian Process Regression for Data Fulfilling Linear Differential Equations with Localized Sources

Abstract

Specialized Gaussian process regression is presented for data that are known to fulfill a given linear differential equation with vanishing or localized sources. The method allows estimation of system parameters as well as strength and location of point sources. It is applicable to a wide range of data from measurement and simulation. The underlying principle is the well-known invariance of the Gaussian probability distribution under linear operators, in particular differentiation. In contrast to approaches with a generic covariance function/kernel, we restrict the Gaussian process to generate only solutions of the homogeneous part of the differential equation. This requires specialized kernels with a direct correspondence of certain kernel hyperparameters to parameters in the underlying equation and leads to more reliable regression results with less training data. Inhomogeneous contributions from linear superposition of point sources are treated via a linear model over fundamental solutions. Maximum likelihood estimates for hyperparameters and source positions are obtained by nonlinear optimization. For differential equations representing laws of physics the present approach generates only physically possible solutions, and estimated hyperparameters represent physical properties. After a general derivation, modeling of source-free data and parameter estimation is demonstrated for Laplace’s equation and the heat/diffusion equation. Finally, the Helmholtz equation with point sources is treated, representing scalar wave data such as acoustic pressure in the frequency domain.

A Survey of Constrained Gaussian Process Regression: Approaches and Implementation Challenges

“…In this sense, the approach is similar to that of Section 6.2, although the kernel is not obtained from a transformation of a prior kernel but rather is constructed from solutions to the problem Lu = 0. Albert and Rath (2020) show that such a covariance kernel must satisfy…”

“…They point out that convolution kernels can also be considered. Albert and Rath (2020) study the Laplace, heat, and Hemholtz equations, performing MLE and inferring the solution u from the PDE constraint and scattered observations. They construct kernels of the form E41, which satisfy Eq.…”

Dependent state space Student‐t processes for imputation and data augmentation in plasma diagnostics

“…[5] In regression with GPs, the posterior distribution of unknown values is inferred, making it a powerful tool for nonlinear multivariate interpolation. [6] Correlations over time are captured by a covariance function or kernel, which may be designed for specific tasks (e.g., to fulfill laws of physics [7,8] ). A generalization of GPs are TPs [9] allowing a more robust estimation of mean and variance for data with outliers.…”

Gaussian Process Regression constrained by Boundary Value Problems

“…While Graepel [10] considered boundary conditions together with differential equation constraints, the approach involved performing regression for a factorized representation of the solution which was constructed in special domains and for Dirichlet conditions; a straightforward construction for general domains was not provided, nor was co-kriging considered. Other related work include those of Owhadi [12], who considered Bayesian numerical homogenization, and Albert and Rath [13], who utilized covariance kernels in the form of a Mercer expansion to enforce PDE constraints. In addition to our unique framework, we go beyond the examples of Solin and Kok [11] by considering general mixed boundary conditions, such as Dirichlet conditions in certain regions of ∂Ω and Neumann conditions in other regions.…”