Physical and non-physical properties (38 features, many of them are irrelevant) of dense pockets of gas (called dense cores) extracted from 5 simulations of colliding molecular clouds from Sakre et al. 2022 (in prep.)
Purpose: 1. To predict model from test cores (50 percent data) 2. To find importance of features (properties)
Method: 1. kNNeighbors (No. of neighbors = 5(default)) is used 2. Random Forest classifier (with max_depth=2) is used to judge feature importance
Results: 1. Found very low accuracy of 30 %, implying properties have weak relation with their model type 2. RFClassifier gave high feature importance to "x-position" feature and low importance to "y-velocity"
Does the Result make sense?
Yes.
1.Low Accuracy: The low accuracy was expected since many of the physical conditions are common and thus it is hard to classify.
2.1. x-position high feature importance: Based on collision speed and cloud size, the x-position can be highly
simulation model dependent.
2.2. y-velocity high feature importance: We expect no specific preffered y-velocity value due to near symmetry of all our 5 collision models along y = 0 plane