Adaptive deep learning for time-varying systems with hidden parameters: Predicting changing input beam distributions of compact particle accelerators

Machine learning (ML) tools such as encoder-decoder deep convolutional neural networks (CNN) are able to extract relationships between inputs and outputs of large complex systems directly from raw data. For time-varying systems the predictive capabilities of ML tools degrade as the systems are no longer accurately represented by the data sets with which the ML models were trained. Re-training is possible, but only if the changes are slow and if new input-output training data measurements can be made online non-invasively. In this work we present an approach to deep learning for time-varying systems in which adaptive feedback based only on available system output measurements is applied to encoded low-dimensional dense layers of encoder-decoder type CNNs. We demonstrate our method in developing an inverse model of a complex charged particle accelerator system, mapping output beam measurements to input beam distributions while both the accelerator components and the unknown input beam distribution quickly vary with time. We demonstrate our results using experimental measurements of the input and output beam distributions of the HiRES ultra-fast electron diffraction (UED) microscopy beam line at Lawrence Berkeley National Laboratory. We show how our method can be used to aid both physics and ML-based surrogate online models to provide non-invasive beam diagnostics and we also demonstrate how our method can be used to automatically track the time varying quantum efficiency map of a particle accelerator's photocathode.