LCLS Cu Injector NN Model

This repository contains the model files corresponding to the LCLS Cu injector NN surrogate model, and example notebooks illustrating how to load and use the model. Using LUME-model is recommended.

Model Description

The model was trained on IMPACT-T simulation data by Auralee Edelen to predict beam properties at OTR2 using injector PVs for LCLS. As the model was trained with normalized data, input and output transformations have to be applied to use it on simulation data. Another layer of transformations is required for using it with EPICS data. See provided examples for more information.

The model YAML provided in this repo handles the full PV to sim to model transformations.

Dependencies

lume-model

Usage

From the main repository directory, call

from lume_model.models import TorchModel

# load model from yaml
model = TorchModel("model_config.yaml")

# evaluate the model at a given point
print(model.evaluate({"QUAD:IN20:425:BACT": -1}))

# get model input variables
print(model.input_variables)

# get model output variables
print(model.output_variables)

NOTE: when not specified, input variables are set to their default values as defined in model_config.yaml

Examples

• Load and print model information
• Load as LUME-model for use with EPICS data

Default Input Variables

The default value for QE01:b1_gradient in the simulation variable specification has been noticed to lie outside the given value range (likewise for QUAD:IN20:425:BACT in the PV variable specification).
Thus, a new value was determined by minimizing the model prediction of the transverse beam size within the valid range (documented in this notebook).

Notes about Working with EPICS PV Values

Unmeasured Input PVs

Some of the input features used as features of the model are not available in EPICS. These include:

distgen:t_dist:length:value

This is the pulse length within the simulation. There has been some discussion about creating a PV to record the pulse length but for now, a reference value of 1.8550514181818183 (PV units) or 3.06083484 (sim units) is used by default.

L0B_scale:voltage

As demonstrated by the train input min and max values in model.json, this value was treated as a constant when training the surrogate model. However in reality, its PV value ACCL:IN20:400:L0B_ADES shows a distribution of values. If it was to be used in the model for predictions, the error would increase dramatically and therefore any measured values from EPICS are overwritten by the value seen during training, scaled to PV units.

distgen:total_charge:value

As above, the charge value was constant in the training dataset but its PV value FBCK:BCI0:1:CHRG_S shows a distribution of values. Measured values from EPICS are overwritten by the value seen during training, scaled to PV units.

distgen:r_dist:sigma_xy:value

The value for the beam size (r_dist) is not measured directly in EPICS but we do measure the XRMS and YRMS value of the beam. We use these PVs (CAMR:IN20:186:XRMS and CAMR:IN20:186:YRMS) to calculate a value for the beam size using the formula:

r_dist = np.sqrt(data["CAMR:IN20:186:XRMS"].values ** 2 + data["CAMR:IN20:186:YRMS"].values ** 2)

We call this computed PV CAMR:IN20:186:R_DIST. Therefore, when pulling data from the archive, this step needs to be completed in any data processing.