leonardoFiedler / neural-network-geoscience

Neural Network And Geoscience

The objective of this study is to use Neural Network`s with PyTorch to classify over 12 classes a rock type.

Types of Rock

Based on column FORCE_2020_LITHOFACIES_LITHOLOGY

• 30000: Sandstone

• 65030: Sandstone/Shale

• 65000: Shale

• 80000: Marl

• 74000: Dolomite

• 70000: Limestone

• 70032: Chalk

• 88000: Halite

• 86000: Anhydrite

• 99000: Tuff

• 90000: Coal

• 93000: Basement

How to Execute the project

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Study

The study is dividided in:

1. Analyse data
2. Baseline With Decision Tree Classifier
3. Different Neural Network Architectures test
4. Generate predictions of a hidden file

Analyse Data and Baseline With Decision Tree Classifier

The analysis and Decision Tree can be found on Base_Analyze.ipynb. This file load data and analyse based on Decision Tree algorithm with unbalanced and balanced Data.

In Unbalanced data, the result was 92.35% and with balanced data with SMOTE (Sinthetic Minotiry Oversampling Technique) was 98.88%.

Different Neural Network Architectures test

The table below describe the tests made with Neural Network. The objective was to start with just few data and add more or change model according to the results.

Number	Model Name	Max Num Samples Per Class	SMOTE?	Num Epochs	Train Acc	Train Loss	Val Acc	Val Loss	Test Precision	Test Recall	Test F1-Score
1	single_layer	10000	False	100	77.390%	0.65312	74.709%	0.68731	0.79	0.75	0.76
2	multi_layer	10000	False	100	72.729%	0.84311	72.398%	0.75672	0.76	0.72	0.73
3	multi_layer	50000	False	1000	59.464%	1.14423	64.466%	1.02543	0.67	0.64	0.63
4	multi_layer_relu	10000	False	100	88.023%	0.39305	87.399%	0.40415	0.88	0.87	0.87
5	multi_layer_relu_batch_norm	10000	False	100	92.029%	0.21409	91.749%	0.19964	0.92	0.92	0.92
6	multi_layer_relu_batch_norm	10000	True	100	93.801%	0.17439	94.958%	0.13612	0.95	0.95	0.95
7	multi_layer_3_relu_batch_norm_dropout	20000	True	300	69.476%	0.75197	72.979%	0.60935	0.74	0.73	0.72

Table Definition

1. Number: Test Number

2. Model Name: Name of the model

3. Max Num Samples Per Class: The maximum number of classes per class for all, train/valid/test

4. SMOTE? Boolean that indicates if the test used augmented data or not

5. Num Epochs: Number of epochs

6. Train Acc: Last epoch train accuracy

7. Train Loss: Last epoch train loss

8. Val Acc: Last epoch valid accuracy

9. Val Loss: Last epoch valid loss

10. Test Precision: Weighted test precision

11. Test Recall: Weighted test recall

12. Test F1-Score: Weighted test F1-Score

Important: Some of models were exectued more than once and some values may vary a little.

Generate predictions of a hidden file

As the result above, we save 2 models in data/model folder:

1. Test 5 with 92% of test precision (unbalanced data)
2. Test 6 with 95% of test precision (balanced data)

The inferences were executed in Inference_Model.ipynb file. The result generates are in data/output folder.

Conclusion

In this homework I could learn more about PyTorch, unbalanced and balanced data and multiple kind of test. Unfortunatelly, for a higher results, more data should be added in some classes to balance and with more analysis on each variable, some of them could be removed (or outliers and others).