Evaluating Neural Network Architectures and Signal Processing Techniques for Diagnostic Reconstruction in DIII-D

Within complex systems that cannot be described analytically, neural network (NN) models have demonstrated the capability to use inter-diagnostic correlations to reconstruct diagnostic data based on an independent set of other diagnostic data. Based on the example of reconstructing interferometer data from Electron Cyclotron Emission (ECE) data in the DIII-D tokamak, we give insights in some crucial design choices to develop efficient and transparent machine learning (ML) pipelines. In particular, the importance of selecting the appropriate signal representation and NN architecture are shown to play a crucial role in the effectiveness of such models.

We compare the impact of different ML models, including recurrent NNs and convolutional NNs. Additionally, we investigate the necessity of including a feature extraction pipeline, consisting of FFT and subsequent filters in the preprocessing step, versus working directly with the raw time-series data. The outcomes of this study can be generalized to other diagnostics, representing a significant step towards implementing efficient signal processing pipelines and machine learning models at DIII-D. This work is a step towards developing a model capable of learning latent features from a multimodal set of diagnostics, thereby enabling the reconstruction of missing diagnostic data.