# ExoMiner: A Highly Accurate and Explainable Deep Learning Classifier to Mine Exoplanets

**Authors:** Hamed Valizadegan, Miguel Martinho, Laurent S. Wilkens, Jon M. Jenkins, Jeffrey Smith, Douglas A. Caldwell, Joseph D. Twicken, Pedro C. Gerum, Nikash Walia, Kaylie Hausknecht, Noa Y. Lubin, Stephen T. Bryson, Nikunj C. Oza  
**arXiv:** [2111.10009](https://arxiv.org/abs/2111.10009)  
**Date:** 2021-11-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.10009v1/

## Key points

- Kepler and TESS missions have generated over 100,000 potential transit signals that must be processed to create a catalog of planet candidates
- Machine learning is being used to analyze this data in search of new exoplanets
- Existing machine learning works do not mimic how domain experts examine diagnostic tests to vet a transit signal
- ExoMiner is a highly accurate, explainable, and robust deep learning classifier that mimics the manual vetting process used by experts
- ExoMiner allows for the validation of 301 new exoplanets from the MAST Kepler Archive and can be applied across missions such as TESS
- The modular design of ExoMiner favors its explainability with a simple framework providing feedback on why ExoMiner classifies a transit signal into a specific class label (e.g., planet candidate or not planet candidate)
- Diagnostic tests are the main components of the 1-page DV summary and longer DV report used by experts to vet transit signals.
- Future directions include changes in architecture to include new important features; performing a full explainability study utilizing more advanced techniques; analyzing label noise; exploring underrepresented scenarios in more depth; adapting ExoMiner for TESS data; providing feedback on why ExoMiner classifies signals into specific classes; incorporating stellar parameters; examining unfolded flux data; studying phase-folded full orbit views; assessing secondary eclipses; investigating whitened transits views; checking odd/even views for false positives due to background objects; analyzing model fits/DV diagnostics parameters etc.

## Summary

The Kepler and TESS missions have generated over 100,000 potential transit signals that must be processed in order to create a catalog of planet candidates. In recent years, there has been a growing interest in using machine learning to analyze this data in search of new exoplanets. However, existing machine learning works do not mimic how domain experts examine diagnostic tests to vet a transit signal. This is where ExoMiner comes into play - it is a highly accurate, explainable, and robust deep learning classifier that mimics the manual vetting process used by experts. ExoMiner allows for the validation of 301 new exoplanets from the MAST Kepler Archive and is general enough to be applied across missions such as the ongoing TESS mission. The modular design of ExoMiner favors its explainability. A simple explainability framework provides experts with feedback on why ExoMiner classifies a transit signal into a specific class label (e.g., planet candidate or not planet candidate). Combined with flux data in multiple forms, these diagnostic tests are the main components of the 1-page DV summary and the longer DV report used by experts to vet transit signals. The 1-page DV summary report includes stellar parameters, unfolded flux data, phase-folded full-orbit and transit view of the flux data, phase-folded transit view of the secondary eclipse, phase-folded whitened transit view, phase-folded odd and even transit view, difference image to check for false positives due to a background object being the source of the transit signature, and an analysis table that provides values related to the model fit and various DV diagnostic parameters. The analysis performed in this paper will allow for future directions for improving and building upon this work. First, changes will be made to the ExoMiner architecture to include new important features such as the number of transits and uncertainty about the depth of odd and even views to ensure that the model's use of these branches is more effective. Second, a full explainability study will be performed by utilizing more advanced techniques and designing a framework that can provide experts with insight about ExoMiner's labels and scores. Thirdly there will be more depth analysis of how label noise affects underrepresented scenarios. In conclusion, ExoMiner is a highly accurate and explainable deep learning classifier that mimics how domain experts examine diagnostic tests to vet a transit signal. It allows for validation of new exoplanets from Kepler Archive and can be applied across missions such as TESS. Future directions include changes in architecture to include new important features; performing a full explainability study utilizing more advanced techniques; analyzing label noise; exploring underrepresented scenarios in more depth; adapting ExoMiner for TESS data; providing feedback on why ExoMiner classifies signals into specific classes; incorporating stellar parameters; examining unfolded flux data; studying phase-folded full orbit views; assessing secondary eclipses; investigating whitened transits views; checking odd/even views for false positives due to background objects; analyzing model fits/DV diagnostics parameters etc..