jlubbersgeo/lubbers23_gcubed

Probabilistic source classification of large tephra producing eruptions using supervised machine learning: An example from the Alaska-Aleutian arc

Jordan Lubbers¹, Matthew Loewen¹, Kristi Wallace¹, Michelle Coombs¹, Jason Addison²
¹U.S. Geological Survey Alaska Volcano Observatory
²U.S. Geological Survey Minerals Energy and Geophysics Science Center

Contact: jlubbers@usgs.gov

This repository is home to the python scripts that recreate the data and figures of the above manuscript. Below is a brief explanation of each notebook.

How to use

Every script in the manuscript is designed to be run using the supplementary spreadsheet associated with the manuscript. Running the scripts can be done using whatever python IDE you choose, however, the user should be somewhat familiar with their terminal/command prompt (or at least how to open and type in it), as the code will prompt the user to enter the filepaths for:

1. where to export the figures/spreadsheets/outputs of the code (this is a folder directory)
   1. ex: "C:\Users\username\...\Gcubed_ML_Manuscript\code_outputs"
2. the path to the data to be used. This is usually one of two things:
   1. The supplementary spreadsheet itself, ex: "C:\Users\username\...\Gcubed_ML_Manuscript\supplementary_spreadsheet.xlsx"
   2. A product derived from previous code, in which case a path to a directory of code outputs is suggested, ex: "C:\Users\username\...\Gcubed_ML_Manuscript\code_outputs"

Code virtual environment

We have created a virtual environment contained within lubbers23gcubed.yml. To create a virtual environment using this file, please see the Anaconda documentation. In brief, from your base environment:

conda env create -f lubbers23gcubed.yml

To double check that it was created:

conda info --envs

And the virtual environment should be in the displayed list. To activate it:

conda activate lubbers23gcubed

Now to run the scripts from the terminal window, simply navigate to the directory where they are saved on your local machine (example using the first script):

cd path\to\downloaded_scripts
python 1_build_training_data.py

Explanation of code notebooks

1. build_training_data.py: this takes the major_trace_proximal_train_data sheet in the supplementary data and builds the training dataset for subsequent use in machine learning applications. To do this, we will largely be using functions from the kinumaax.crunching module to clean the data and the pyrolite package to transform the data. It also cleans and transforms data in the iodp_trace_data, and derkachev_data sheets such that the eventual machine learning model can be applied to them.

2. bivariate_plots.py: This notebook is devoted to creating Figure 2.

3. model_feature_selcetion.py: This notebook explores the feature engineering process whereby a basic random forest model is trained using a variety of features such that the results from each feature space can be compared, ultimately producing Figures 3, A3, A4.

4. gradient_boosting_tuning.py: Using a grid search to find the optimal parameters for the gradient boosting classifier used in the final ensemble voting classifier. Produces Figure A7B

5. KNN_tuning.py: Using a grid search to find the optimal parameters for the K-nearest neighbors classifier used in the final ensemble voting classifier. Produces Figure A6A.

6. LDA_tuning.py: Using a grid search to find the optimal parameters for the Linear Discriminant Analysis classifier used in the final ensemble voting classifier. Produces Figure A5A.

7. logistic_regression_tuning.py: Using a grid search to find the optimal parameters for the Logistic Regression classifier used in the final ensemble voting classifier. Produces Figure A5B.

8. random_forest_tuning.py: Using a grid search to find the optimal parameters for the Random Forest classifier used in the final ensemble voting classifier. Produces Figure A7A.

9. SVM_tuning.py: Using a grid search to find the optimal parameters for the Support Vector Machine classifier used in the final ensemble voting classifier. Produces Figure A6B.

10. final_voting_classifier.py: Builds thet final ensemble voting classifier using the optimal hyperparameters for each individual algorithm. The influence of each individual aglorithm is determined by its F₁ score. Produces Figures 5 and A8.

11. random_forest_proximities.py: Uses an out of the box random forest classifier and the optimal features found in model_feature_selection.py to compare class similarities as viewed by a random forest algorithm. Produces Figure 4.

12. variance_analysis.py: completes the variance test as described the section "Assessing model variance" where we test how likely we would be able to correctly predict thte volcanic source for an eruption if the model was not trained on that eruption.

13. variance_analysis_plots.py: Visualizing the results of the code run in variance_analysis_plots.py. Produces Figures 6, 7, 8.

14. IODP_novelty_detection.py: Makes plots in multivariate space to compare where our training data plots with respect to IODP data to see if they overlap. Produces Figures 9, A9, A10.

15. IODP_predictions.py: Uses the final model developed in final_voting_classifier.py to predict the most likely volcanic source for each unknown sample listed in Table A4. Produces Figures 10, 11, 12.

16. derkachev_predictions.py: Uses the final model developed in final_voting_classifier.py to predict the most likely volcanic source for samples Br2 and SR2 from Derkachev et al., (2018). Produces Figures 13, 14.

17. EPMA_stds_check.py: Quantifies the accuracy and precision of EPMA secondary standards analyzed in the study. Produces values in Table A2 and Figure A1.

18. lasercalc_secondary_standards.py: Determines concentrations of LA-ICP-MS secondary standards analyzed in this study. Input data are secondary standards data normalized to ²⁹Si using LaserTRAM-DB. Produces values in Table A3 and Figure A2.

19. aleutian_colors.py: creates a dictionary of volcano_name : {**kwargs} pairs for each volcano in the dataset such that it can be consistently plotted in all figures.

20. mpl_defaults.py: establishes default matplotlib settings and helper functions to produce figures for the manuscript.

21. kinumaax: A small collection of python scripts to help working with tephra data

    1. crunching.py: has functions used for data cleaning, filtering, unit conversions, etc. Largely used by build_training_data.py

    2. learning.py: builds a machine learning pipeline for volcano classification using tephra geochemistry data. Inherits BaseEstimator and ClassifierMixin base classes from scikit-learn. This will store the data, performance metrics, and predictions in an object to be accessed as attributes and functions can be called as methods. Used largely in model_feature_selection.py to efficiently create numerous Random Forest classifier models.

    3. visualization.py: various functions for producing plots molre specific to volcanology/petrology. Not used in this manuscript.