Graph neural networks for particle accelerator facilities
A method or tool for an efficient, data-driven approach to beam tuning in particle accelerators that leverages deep learning over structured data. A beamline is represented as a graph, where individual elements are nodes and relationships between elements are edges. Element parameters are captured as node features. A graph neural network is used to generate a whole-graph embedding that preserves structural information, and dimensionality reduction techniques are applied to visualize low-dimensional representations. Providing a category label for each embedding to identify optimal regions of parameter space. The method serves as a global diagnostic, inasmuch as it monitors a high-dimensional space and provides feedback to operators when changes are made. Operators can track if changes during beam tuning move the configuration away or towards optimal regions of parameter space. On-line and off-line tool embodiments of the beam tuning tool are described.
1. A method of representing the state of a particle accelerator beamline with a low-dimensional, embedding, the method comprising:
representing a beamline as a graph, with nodes representing elements, node features representing element attributes, and edges used to capture relationships between elements;
pre-training a graph neural network in a self-supervised way using unlabeled, historical operational data;
fine-tuning the graph neural network on a small set of labeled data in a supervised way;
implementing scheduled training to maintain performance of the method, whereby the method is re-trained at regular intervals using data collected since the previous scheduled training;
visualizing the graph embeddings generated by the trained model on new data by reducing the dimensionality to two or three dimensions for ease of interpretation;
monitoring the stability of the particle accelerator beamline, wherein monitoring the stability of the particle accelerator comprises updating the beamline graph embedding and visualization in short, regular intervals; and observing the level of jitter in parameter space;
providing a means for operators to define a bounding box for a desired region of parameter space; and
providing an alert system for notifying when the system exceeds the boundaries of the bounding box.
2. The method of claim 1 comprises exploring a high-dimensional space representing historical operational data over man′ months or years identifying trends and patterns in the data.
3. The method of claim 1 comprising analyzing attention weights of the graph neural network to provide a metric for the importance of the relationships between nodes, said relationships revealing complex relationships between beamline components.
4. The method of claim 1 comprising:
identifying optimal regions of parameter space for graph embeddings; and
labeling good and bad subsets of the graph embeddings.
5. The method of claim 4 comprising:
an off-line tool for rapid beamline optimization; and
said nodes including setting nodes and reading nodes.
6. The method of claim 5 wherein said off-line tool comprises suppressing the reading nodes of the graph representation of the beamline.
7. The method of claim 6 comprising:
enabling users of said off-line tool to modify said setting nodes in an off-line manner; and
converting the new configuration after each modification to a graph input to a pre-trained model; and
plotting the embedding to display the effect of the change of settings, said plot displaying the state of the beamline with respect to good regions of the parameter space.
8. The method of claim 1 comprising an on-line tool to provide near real-time feedback on the effect of tuning changes.
9. The method of claim 8 wherein said online tool comprises:
tracking in real-time how a beamline setup is moving through the low-dimensional space during beam tuning; and
determining if operating changes are moving the system closer or farther from optimal regions.