Classify events based on brem or not 2 classes of events Brem, with recoil energy < 1500 MeV No brem See if we can distinguish them, two cases should have different shower features Compare 3D CNN, Graph Neural Network 100 000 events for training+ validation 100 000 events for testing/predictions 1
3D CNN Architecture 1 conv layer, 32 filters of size (8,8,8) 1 max pool of size (5,5,5) 1 dense layer Dropout 0.16 Adam optimizer with learning rate 0.004 ReLu Activation function after conv and dense layer Max(0,x) Input size = (45,70,35) 96.4 % accuracy 2
Graph Neural Network Architecure 3 EdgeConv blocks 1 deep layer 256 nodes Dropout 0.1 Adam optimizer with learning 0.003 100 hits as input vertices ReLu activation function 83.0% accuracy 3
ROC curves, CNN & Graph 4
Mean and Std Dev 5
Mean and Std Dev Rates 6
Mean and Std Dev Rates 7
Mean and Std Dev 8
Mean and Std Dev 9
Mean and Std Dev 10
Energy 11
Azimuthal Angle 12
Azimuthal Angle 13
Energy vs Hits 14
Energy & Hits 15
Energy vs Angle 16
Energy vs Angle 17
Recommend
More recommend