An adaptive approach to machine learning for compact particle accelerators.

Area of Science:

• Physics
• Computer Science
• Materials Science

Background:

• Machine learning (ML) models struggle with time-varying systems where data drifts over time.
• Retraining ML models for complex systems is often infeasible due to slow data acquisition rates.
• Existing ML approaches require extensive retraining when system dynamics change.

Purpose of the Study:

• To develop a deep learning method for time-varying systems that adapts without retraining.
• To create an inverse model for complex accelerator systems using adaptive deep convolutional neural networks (CNNs).
• To enable non-invasive diagnostics for systems with rapidly changing parameters.

Main Methods:

• Implemented an adaptive feedback mechanism within the encoder-decoder CNN architecture.
• Applied feedback in low-dimensional dense layers, using only available system output measurements.
• Developed an inverse model to map output beam measurements to input beam distributions for accelerator systems.

Main Results:

• Successfully demonstrated the adaptive CNN approach on experimental data from the HiRES ultra-fast electron diffraction (UED) beam line.
• Showcased automatic tracking of time-varying photocathode quantum efficiency maps.
• Validated the method's ability to handle rapidly changing accelerator components and input beam distributions.

Conclusions:

• The proposed adaptive deep learning method effectively addresses challenges in modeling time-varying systems.
• This approach eliminates the need for continuous retraining, improving efficiency and applicability.
• The technique offers a powerful tool for non-invasive beam diagnostics in complex experimental setups.