Bayesian uncertainty calculation in neural network inference of ion and electron temperature profiles at W7-X

Aleatoric Uncertainty for Errors-in-Variables Models in Deep Regression

A Bayesian treatment of deep learning allows for the computation of uncertainties associated with the predictions of deep neural networks. We show how the concept of Errors-in-Variables can be used in Bayesian deep regression to also account for the uncertainty associated with the input of the employed neural network. The presented approach thereby exploits a relevant, but generally overlooked, source of uncertainty and yields a decomposition of the predictive uncertainty into an aleatoric and epistemic part that is more complete and, in many cases, more consistent from a statistical perspective. We discuss the approach along various simulated and real examples and observe that using an Errors-in-Variables model leads to an increase in the uncertainty while preserving the prediction performance of models without Errors-in-Variables. For examples with known regression function we observe that this ground truth is substantially better covered by the Errors-in-Variables model, indicating that the presented approach leads to a more reliable uncertainty estimation.

Neural network surrogates of Bayesian diagnostic models for fast inference of plasma parameters

We present a framework for training artificial neural networks (ANNs) as surrogate Bayesian models for the inference of plasma parameters from diagnostic data collected at nuclear fusion experiments, with the purpose of providing a fast approximation of conventional Bayesian inference. Because of the complexity of the models involved, conventional Bayesian inference can require tens of minutes for analyzing one single measurement, while hundreds of thousands can be collected during a single plasma discharge. The ANN surrogates can reduce the analysis time down to tens/hundreds of microseconds per single measurement. The core idea is to generate the training data by sampling them from the joint probability distribution of the parameters and observations of the original Bayesian model. The network can be trained to learn the reconstruction of plasma parameters from observations and the model joint probability distribution from plasma parameters and observations. Previous work has validated the application of such a framework to the former case at the Wendelstein 7-X and Joint European Torus experiments. Here, we first give a description of the general methodological principles allowing us to generate the training data, and then we show an example application of the reconstruction of the joint probability distribution of an effective ion charge Z_eff-bremsstrahlung model from data collected at the latest W7-X experimental campaign. One key feature of such an approach is that the network is trained exclusively on data generated with the Bayesian model, requiring no experimental data. This allows us to replicate the training scheme and generate fast, surrogate ANNs for any validated Bayesian diagnostic model.

Deep learning for Gaussian process soft x-ray tomography model selection in the ASDEX Upgrade tokamak

Gaussian process tomography (GPT) is a method used for obtaining real-time tomographic reconstructions of the plasma emissivity profile in tokamaks, given some model for the underlying physical processes involved. GPT can also be used, thanks to Bayesian formalism, to perform model selection, i.e., comparing different models and choosing the one with maximum evidence. However, the computations involved in this particular step may become slow for data with high dimensionality, especially when comparing the evidence for many different models. Using measurements collected by the Soft X-Ray (SXR) diagnostic in the ASDEX Upgrade tokamak, we train a convolutional neural network to map SXR tomographic projections to the corresponding GPT model whose evidence is highest. We then compare the network's results, and the time required to calculate them, with those obtained through analytical Bayesian formalism. In addition, we use the network's classifications to produce tomographic reconstructions of the plasma emissivity profile.

Bayesian modeling of microwave radiometer calibration on the example of the Wendelstein 7-X electron cyclotron emission diagnostic

This paper reports about a novel approach to the absolute intensity calibration of an electron cyclotron emission (ECE) spectroscopy system. Typically, an ECE radiometer consists of tens of separated frequency channels corresponding to different plasma locations. An absolute calibration of the overall diagnostic including near plasma optics and transmission line is achieved with blackbody sources at LN₂ temperature and room temperature via a hot/cold calibration mirror unit. As the thermal emission of the calibration source is typically a few thousand times lower than the receiver noise temperature, coherent averaging over several hours is required to get a sufficient signal to noise ratio. A forward model suitable for any radiometer calibration using the hot/cold method and a periodic switch between them has been developed and used to extract the voltage difference between the hot and cold temperature source via Bayesian analysis. In contrast to the classical analysis which evaluates only the reference temperatures, the forward model takes into account intermediate effective temperatures caused by the finite beam width and thus uses all available data optimally. This allows the evaluation of weak channels where a classical analysis would not be feasible, is statistically rigorous, and provides a measurement of the beam width. By using a variance scaling factor, a model sensitive adaptation of the absolute uncertainties can be implemented, which will be used for the combined diagnostic Bayesian modeling analysis.