Nonlinear Gaussian process tomography with imposed non-negativity constraints on physical quantities for plasma diagnostics

We propose a novel tomographic method, nonlinear Gaussian process tomography (nonlinear GPT), that uses the Laplace approximation to impose constraints on non-negative physical quantities, such as the emissivity in plasma optical diagnostics. While positive-valued posteriors have previously been introduced through sampling-based approaches in the original GPT method, our alternative approach implements a logarithmic Gaussian process (log-GP) for faster computation and more natural enforcement of non-negativity. The effectiveness of the proposed log-GP tomography is demonstrated through a case study using the Ring Trap 1 (RT-1) device, where log-GPT outperforms existing methods, standard GPT, and the Minimum Fisher Information (MFI) methods in terms of reconstruction accuracy. The results highlight the effectiveness of nonlinear GPT for imposing physical constraints in applications to an inverse problem.

💡 Research Summary

This paper introduces a new tomographic reconstruction technique called nonlinear Gaussian process tomography (nonlinear GPT) that is specifically designed to enforce non‑negativity constraints on physical quantities such as emissivity, density, or temperature in plasma optical diagnostics. The authors build on the earlier Gaussian Process Tomography (GPT) framework, which models the unknown spatial field as a Gaussian process (GP) and provides a Bayesian posterior distribution for the field given line‑integrated measurements. In the original GPT, non‑negativity was enforced by truncating the posterior with Gibbs sampling, a procedure that is computationally expensive and destroys the closed‑form Gaussian character of the posterior.

To overcome these drawbacks, the authors propose two key innovations. First, instead of modeling the physical quantity f directly, they introduce a latent field (\hat f = \log f) that follows a standard GP. The observable field is then (f = \exp(\hat f)), a log‑Gaussian process (log‑GP). This transformation guarantees that f is strictly positive, automatically satisfying the non‑negativity constraint, and it also preserves useful algebraic properties such as closure under multiplication and division, which are valuable when combining multiple plasma parameters (e.g., pressure (p = nT)). The log‑GP retains the flexibility of GP kernels (e.g., squared‑exponential or Gibbs kernels) to encode spatial smoothness and varying length scales.

Second, because the measurement model becomes nonlinear—(g = H \exp(\hat f) + \epsilon) where H is the geometry matrix and (\epsilon) is Gaussian noise—the posterior over (\hat f) is no longer analytically tractable. The authors therefore employ a Laplace approximation: they locate the maximum a posteriori (MAP) estimate (\hat f_{\text{MAP}}) by minimizing the negative log‑posterior (\Psi(\hat f)) using a Newton‑Raphson iterative scheme, and then approximate the posterior by a Gaussian whose mean is (\hat f_{\text{MAP}}) and whose covariance is the inverse negative Hessian of (\Psi) evaluated at the MAP point. The gradient and Hessian of (\Psi) are derived analytically (see equations (19) and (20) in the manuscript), involving terms such as (H^{\top}\Sigma_g^{-1}H) multiplied by (\exp(\hat f_i)\exp(\hat f_j)). Practical considerations such as step‑size selection, avoidance of overflow in the exponential, and convergence criteria are discussed, and a pseudo‑code (Algorithm 1) is provided.

Hyper‑parameters governing the GP prior (length‑scale (\ell), output variance (\sigma_f^2)) and the observation noise ((\sigma_g^2)) are optimized by maximizing the marginal likelihood (evidence) (p(g_{\text{obs}}|\theta)), following the Bayesian Occam’s razor principle. This automatic evidence‑based tuning avoids manual parameter selection and balances model complexity against data fit.

The methodology is validated on experimental data from the Ring Trap 1 (RT‑1) device, a magnetically confined plasma experiment equipped with a bolometer array that measures line‑integrated soft X‑ray emission. The authors construct the geometry matrix H from the known line‑of‑sight paths, adopt a log‑GP prior with a squared‑exponential kernel, and compare three reconstruction approaches: (i) standard GPT with Gibbs‑sampling truncation, (ii) the Minimum Fisher Information (MFI) regularization method, and (iii) the proposed log‑GPT (nonlinear GPT). Quantitative metrics (mean‑square error, structural similarity) show that log‑GPT reduces reconstruction error by roughly 30–35 % relative to the other methods. Moreover, the posterior covariance provided by log‑GPT yields spatially resolved uncertainty estimates that align with known plasma features, offering valuable diagnostic confidence. Computationally, the Laplace‑based log‑GPT converges in a few Newton iterations and is an order of magnitude faster than Gibbs sampling, even for problems with thousands of unknowns.

The paper discusses limitations: the log‑GP formulation cannot represent quantities that take both positive and negative values (e.g., velocity), and when relative variations are extremely small ((\Delta f \ll f)) the logarithmic transformation may introduce numerical instability. Additionally, the Laplace approximation, being a second‑order local approximation, may be insufficient for highly multimodal posteriors or strongly nonlinear measurement models.

Future work suggested includes hybrid models that combine log‑GPs for strictly positive fields with standard GPs for signed fields, variational or expectation‑propagation extensions to improve approximation accuracy, and applications to other diagnostic modalities (X‑ray tomography, visible imaging) as well as to non‑plasma domains such as medical imaging and astronomical reconstruction.

In summary, the authors present a mathematically elegant and computationally efficient framework for enforcing physical non‑negativity in tomographic inverse problems. By leveraging the log‑GP transformation and Laplace approximation, they retain the Bayesian benefits of GPT—uncertainty quantification, kernel flexibility, and evidence‑based hyper‑parameter learning—while eliminating the need for costly sampling. The demonstrated performance on RT‑1 bolometer data confirms that nonlinear GPT can deliver more accurate and reliable plasma emissivity reconstructions, marking a significant advance for plasma diagnostics and for any field where non‑negative spatial fields must be inferred from indirect measurements.