Molecular Analysis of Stellar Clusters Final Paper

By testing a multitude of models from the python package scikit-learn upon the APOGEE dataset, we’ve managed to produce a resultant product that can predict a star’s iron concentration depending upon its gravitational pull (LOGG) value or effective temperature (TEFF) value.

WIth the diversity of research methods and measurement instruments found in astronomy, it is often difficult to draw direct comparisons between conclusions due to differences in both kind and quality of measurements. By using a machine learning algorithm we have been able to predict measurements that may not have been gathered, notably , the iron content of a star based upon temperature or gravity readings. The problem has not been answered before with this data set, and the scientific community does not have a large amount of research that has pertained to this specific APOGEE data release. By addressing this previously unaddressed problem, we can advance the scientific community with a new contribution that allows scientists of the future to gain additional data in conditions where such data may not be so readily available. We have approached the question by using various machine learning algorithms to discover correlations between the data points we were given, creating a predictive model to produce a numerical result. Gravity and temperature do not present a direct linear relationship to the iron content of a star. However, as their data points tend to cluster together, we can utilize the KNeighborsRegressor to find correlations in the way that these points have, and produce a numerical prediction. Throughout testing we’ve found that the KNeighborsRegressor does better on all accounts in comparison to a large variety of other regression models such as the GaussianProcessRegressor. Considering the results of the Mean Absolute Percentage Error function upon the results of both of these, we can note how the LinearRegressor, mathematically, does marginally better than the KNeighborsRegressor, at a value of 1.13% of average error in comparison to 2.53%.

Explore More!

Research Paper