Frequency-dependent machine learning model in seismic interpretation
Frequency-dependent machine-learning (ML) models can be used to interpret seismic data. A system can apply spectral decomposition to pre-processed training data to generate frequency-dependent training data of two or more frequencies. The system can train two or more ML models using the frequency-dependent training data. Subsequent to training the two or more ML models, the system can apply the two or more ML models to seismic data to generate two or more subterranean feature probability maps. The system can perform an analysis of aleatoric uncertainty on the two or more subterranean feature probability maps to create an uncertainty map for aleatoric uncertainty. Additionally, the system can generate a filtered subterranean feature probability map based on the uncertainty map for aleatoric uncertainty.
1. A method comprising:
applying spectral decomposition to pre-processed training data to generate frequency-dependent training data of two or more frequencies;
training two or more machine-learning (ML) models using the frequency-dependent training data, wherein each ML model of the two or more ML models comprises a plurality of layers, wherein each ML model of the two or more ML models is trained using frequency-dependent training data of a different frequency than a frequency of frequency-dependent training data that is used to train a different ML model of the two or more ML models;
subsequent to training the two or more ML models, applying the two or more ML models to seismic data to generate two or more subterranean feature probability maps;
performing an analysis of aleatoric uncertainty on the two or more subterranean feature probability maps to create an uncertainty map for aleatoric uncertainty; and
generating a filtered subterranean feature probability map based on the uncertainty map for aleatoric uncertainty.
2. The method of claim 1 , wherein performing the analysis of aleatoric uncertainty comprises:
selecting one of the two or more subterranean feature probability maps;
applying an algorithm to quantify aleatoric uncertainty in the selected subterranean feature probability map; and
creating the uncertainty map for aleatoric uncertainty.
3. The method of claim 2 , wherein generating the filtered subterranean feature probability map comprises applying the uncertainty map to filter out high uncertainty areas from the selected subterranean feature probability map.
4. The method of claim 1 , further comprising performing a structure-oriented filter on field seismic data to produce the pre-processed training data.
5. The method of claim 4 , wherein performing the structure-oriented filter comprises optimizing a size of a moving filter window to preserve geologic discontinuities and reduce noise in the field seismic data.
6. The method of claim 1 , wherein the two or more subterranean feature probability maps comprise two or more geologic fault probability maps.
7. The method of claim 1 , wherein training the two or more ML models comprises:
randomly selecting partial training data from the frequency-dependent training data; and
measuring, from the partial training data, a probability distribution function that can quantify uncertainty in the ML model.
8. A system comprising:
a processor; and
a non-transitory computer-readable medium comprising instructions that are executable by the processor for causing the processor to perform operations comprising:
applying spectral decomposition to pre-processed training data to generate frequency-dependent training data of two or more frequencies;
training two or more machine-learning (ML) models using the frequency-dependent training data in which each ML model of the two or more ML models is trained using frequency-dependent training data of a different frequency than a frequency of frequency-dependent training data that is used to train a different ML model of the two or more ML models, wherein each ML model of the two or more ML models comprises a plurality of layers;
subsequent to training the two or more ML models, applying the two or more ML models to seismic data to generate two or more subterranean feature probability maps;
performing an analysis of aleatoric uncertainty on the two or more subterranean feature probability maps to create an uncertainty map for aleatoric uncertainty; and
generating a filtered subterranean feature probability map based on the uncertainty map for aleatoric uncertainty.
9. The system of claim 8 , wherein performing the analysis of aleatoric uncertainty comprises:
selecting one of the two or more subterranean feature probability maps;
applying an algorithm to quantify aleatoric uncertainty in the selected subterranean feature probability map; and
creating the uncertainty map for aleatoric uncertainty.
10. The system of claim 9 , wherein generating the filtered subterranean feature probability map comprises applying the uncertainty map to filter out high uncertainty areas from the selected subterranean feature probability map.
11. The system of claim 8 , wherein the operations further comprise performing a structure-oriented filter on field seismic data to produce the pre-processed training data.
12. The system of claim 11 , wherein performing the structure-oriented filter comprises optimizing a size of a moving filter window to preserve geologic discontinuities and reduce noise in the field seismic data.
13. The system of claim 8 , wherein the two or more subterranean feature probability maps comprise two or more geologic fault probability maps.
14. The system of claim 8 , wherein training the two or more ML models comprises:
randomly selecting partial training data from the frequency-dependent training data; and
measuring, using the partial training data, a probability distribution function that can be used to quantify uncertainty in the ML model.
15. A non-transitory computer-readable medium comprising instructions that are executable by a processing device for causing the processing device to perform operations comprising:
applying spectral decomposition to pre-processed training data to generate frequency-dependent training data of two or more frequencies;
training two or more machine-learning (ML) models using the frequency-dependent training data in which each ML model of the two or more ML models is trained using frequency-dependent training data of a different frequency than a frequency of frequency-dependent training data that is used to train a different ML model of the two or more ML models, wherein each ML model of the two or more ML models comprises a plurality of layers;
subsequent to training the two or more ML models, applying the two or more ML models to seismic data to generate two or more subterranean feature probability maps;
performing an analysis of aleatoric uncertainty on the two or more subterranean feature probability maps to create an uncertainty map for aleatoric uncertainty; and
generating a filtered subterranean feature probability map based on the uncertainty map for aleatoric uncertainty.
16. The non-transitory computer-readable medium of claim 15 , wherein performing the analysis of aleatoric uncertainty comprises:
selecting one of the two or more subterranean feature probability maps;
applying an algorithm to quantify aleatoric uncertainty in the selected subterranean feature probability map; and
creating the uncertainty map for aleatoric uncertainty.
17. The non-transitory computer-readable medium of claim 16 , wherein generating the filtered subterranean feature probability map comprises applying the uncertainty map to filter out high uncertainty areas from the selected subterranean feature probability map.
18. The non-transitory computer-readable medium of claim 15 , wherein the operations further comprise performing a structure-oriented filter on field seismic data to produce the pre-processed training data.
19. The non-transitory computer-readable medium of claim 18 , wherein performing the structure-oriented filter further comprises optimizing a size of a moving filter window to preserve geologic discontinuities and reduce noise in the field seismic data.
20. The non-transitory computer-readable medium of claim 15 , wherein the two or more subterranean feature probability maps comprise two or more geologic fault probability maps.