Analisis Komparatif Model ANN, LSTM, Dan Siamese-LSTM Dalam Estimasi Porositas Berbasis Data Magnetotelurik Pada Sistem Geotermal: Kasus Utah FORGE

Putra, I Gede Nova Darma (2026) Analisis Komparatif Model ANN, LSTM, Dan Siamese-LSTM Dalam Estimasi Porositas Berbasis Data Magnetotelurik Pada Sistem Geotermal: Kasus Utah FORGE. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5017221050-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Indonesia memiliki potensi energi geotermal terbesar kedua di dunia, namun eksplorasi yang efisien masih terkendala oleh keterbatasan data sumur yang mahal dan tidak mencakup area luas. Porositas batuan reservoir mengendalikan kapasitas simpan dan mobilitas fluida hidrotermal, sehingga karakterisasinya menjadi faktor penentu dalam eksplorasi geotermal. Untuk mendukung pengembangan metode estimasi porositas tanpa ketergantungan penuh pada data sumur, data MT dan well log sekunder lapangan Utah FORGE digunakan sebagai data training dan test dalam penelitian ini. Resistivitas hasil inversi magnetotelurik (MT) 2-D diintegrasikan dengan tiga arsitektur machine learning, yaitu Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), dan Siamese-LSTM, pada sumur FORGE 78B-32, Utah, Amerika Serikat. Inversi 2-D menghasilkan model resistivitas dengan normalized RMS sebesar 0,88, yang divalidasi terhadap log AT90 dengan RMSE sebesar 8,9%. Dataset akhir `terdiri atas 1.476 titik beresolusi 1 meter dengan input variabel berupa resistivitas MT dan kedalaman, serta target berupa porositas efektif (NDPHI). Ketiga model dioptimalkan menggunakan Bayesian Optimization dengan pembagian data berbasis kedalaman. Pada data training, ANN mencapai R² = 0,8745 (MAPE = 5,82%) dan LSTM mencapai R² = 0,8658 (MAPE = 5,09%), sementara Siamese-LSTM mencapai R² = 0,5446 (MAPE = 10,42%), mengindikasikan kemungkinan underfitting akibat kapasitas model yang lebih kecil.. Pada data test, LSTM menunjukkan performa terbaik (MAPE = 5,10%; RMSE = 0,0018), diikuti Siamese-LSTM (MAPE = 5,62%) dan ANN (MAPE = 6,22%). Nilai R² negatif pada seluruh data test merupakan konsekuensi matematis dari variansi target yang sempit, diperkuat oleh distributional shift karena interval test berada di luar jangkauan kedalaman domain training. Ketiga model menunjukkan akurasi yang sebanding pada data test, dengan kegagalan generalisasi yang konsisten mengindikasikan keterbatasan utama bersumber dari skema pembagian data, bukan pemilihan arsitektur. Pendekatan ini berpotensi dikembangkan lebih lanjut untuk diaplikasikan pada lapangan panas bumi di Indonesia.
===================================================================================================================================
Indonesia holds the world's second-largest geothermal energy potential, yet efficient exploration remains constrained by the high cost and limited spatial coverage of well data. Reservoir rock porosity directly governs the storage capacity and mobility of hydrothermal fluids, making its characterization a critical factor in geothermal exploration. To support the development of porosity estimation methods that do not rely entirely on well data, magnetotelluric (MT) and secondary well log data from the Utah FORGE field were used as training and testing data in this study. Resistivity derived from 2-D magnetotelluric inversion was integrated with three machine learning architectures, Artificial Neural Network (ANN), Long Short-Term Memory (LSTM), and Siamese-LSTM at well FORGE 78B-32, Utah, United States. The 2-D inversion produced a resistivity model with a normalized RMS of 0.88, validated against the AT90 log with an RMSE of 8.9%. The final dataset consisted of 1,476 points at 1-meter resolution, with MT resistivity and depth as input variables and effective neutron porosity (NDPHI) as the target, optimized using Bayesian Optimization with a depth-based data split. On the training set, ANN achieved R² = 0.8745 (MAPE = 5.82%) and LSTM achieved R² = 0.8658 (MAPE = 5.09%), while Siamese-LSTM achieved R² = 0.5446 (MAPE = 10.42%), indicating possible underfitting due to its smaller model capacity. On the test set, LSTM showed the best performance (MAPE = 5.10%; RMSE = 0.0018), followed by Siamese-LSTM (MAPE = 5.62%) and ANN (MAPE = 6.22%). The negative R² values observed across all test sets are a mathematical consequence of the narrow variance in the target range, compounded by distributional shift, as the test interval lies entirely outside the depth range of the training domain. All three models showed comparable accuracy on the test set, and their consistent generalization failure indicates that the primary limitation stems from the data-splitting scheme rather than architecture choice. This approach has the potential to be further developed for application to geothermal fields in Indonesia.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Magnetotelurik, Porositas, Machine Learning, ANN, LSTM, Siamese-LSTM, Utah FORGE, Geotermal, Magnetotellurics, Porosity, Machine Learning, ANN, LSTM, Siamese-LSTM, Utah FORGE, Geothermal
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QE Geology
Divisions:	Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Geophysics Engineering > 33201-(S1) Undergraduate Thesis
Depositing User:	I Gede Nova Darma Putra
Date Deposited:	23 Jul 2026 09:17
Last Modified:	23 Jul 2026 09:17
URI:	http://repository.its.ac.id/id/eprint/136526

Actions (login required)

View Item