Simulasi Physics Informed Neural Network Untuk Solusi Perpindahan Panas 2D Steady State Pada Material Berpori Dan Metal Organic Framework (MOF-5)

Garibaldi, Rayhan Achmad (2026) Simulasi Physics Informed Neural Network Untuk Solusi Perpindahan Panas 2D Steady State Pada Material Berpori Dan Metal Organic Framework (MOF-5). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5011221067_Undergraduate_Thesis.pdf Restricted to Repository staff only Download (11MB) | Request a copy

Abstract

Material berpori seperti Metal-Organic Framework-5 (MOF-5) memiliki potensi besar dalam aplikasi penyimpanan energi, namun karakterisasi konduktivitas termal efektifnya secara eksperimental sulit dilakukan akibat kompleksitas mikrostruktur porinya. Penelitian ini bertujuan menganalisis performa Physics-Informed Neural Networks (PINN) dalam menyelesaikan persamaan konduksi panas 2D steady-state pada media berpori, serta mengkaji pengaruh variasi geometri pori dan porositas terhadap konduktivitas termal efektif yang diprediksi PINN. Simulasi dilakukan pada 24 konfigurasi yang mencakup dua material (hipotesis dengan ksolid : kvoid = 1,0 : 0,1 dan MOF-5 dengan ksolid = 0,32 W/(m·K), kvoid = 0,024 W/(m·K)), tiga variasi geometri pori (teratur densitas rendah, teratur densitas tinggi, dan acak), serta empat tingkat porositas (30%–60%). Model PINN dikembangkan menggunakan arsitektur fully connected neural network enam lapisan dengan Fourier Feature Embedding dan strategi optimasi hibrid Adam–L-BFGS, divalidasi terhadap solusi finite difference method (FDM) secara numerik serta model Maxwell-Eucken dan Bruggeman secara analitik. Hasil menunjukkan bahwa PINN mampu memprediksi distribusi temperatur dan konduktivitas termal efektif dengan akurasi tinggi, dengan MSE terhadap FDM berkisar 1,97×10⁻⁶–1,43×10⁻⁴ (hipotesis) dan mengalami kenaikan nilai sebesar 1,5-1,7 kali untuk kasus MOF-5. Deviasi terhadap model analitik berkisar 0,50%–1,60% (Maxwell-Eucken, geometri teratur densitas rendah) dan 0,25%–6,12% (Bruggeman, geometri acak), dengan akurasi yang menurun seiring meningkatnya kompleksitas geometri pori. Penelitian ini membuktikan bahwa PINN merupakan pendekatan komputasi yang akurat dan fleksibel untuk analisis termal media berpori heterogen, sekaligus menunjukkan kemampuannya dalam menangkap perilaku fisika yang melampaui prediksi model analitik konvensional.
=================================================================================================================================
Porous materials such as Metal-Organic Framework-5 (MOF-5) hold significant promise for energy storage applications; however, experimental characterization of their effective thermal conductivity is challenging due to the complexity of their pore microstructure. This study aims to analyze the performance of Physics-Informed Neural Networks (PINNs) in solving 2D steady-state heat conduction equations in porous media, and to investigate the influence of pore geometry and porosity variations on the effective thermal conductivity predicted by PINNs. Simulations were conducted across 24 configurations covering two materials (a hypothetical material with ksolid : kvoid = 1.0 : 0.1, and MOF-5 with ksolid = 0.32 W/(m·K) and kvoid = 0.024 W/(m·K)), three pore geometry types (low-density regular, high-density regular, and random), and four porosity levels (30%–60%). The PINN model was developed using a six-layer fully connected neural network architecture with Fourier Feature Embedding and a hybrid Adam–L-BFGS optimization strategy, validated numerically against Finite Difference Method (FDM) solutions and analytically against Maxwell-Eucken and Bruggeman Effective Medium Theory models. Results show that PINNs accurately predict temperature distribution and effective thermal conductivity, with MSE against FDM ranging from 1.97×10⁻⁶ to 1.43×10⁻⁴ (hypothetical material) and NRMSE of 0.22%–3.27% (MOF-5). Deviations from analytical models ranged from 0.50%–1.60% against Maxwell-Eucken (low-density regular geometry) and 0.25%–6.12% against Bruggeman (random geometry), with accuracy decreasing as pore geometry complexity increased. This study demonstrates that PINNs constitute an accurate and flexible computational approach for thermal analysis of heterogeneous porous media, while also revealing their capacity to capture physical behavior beyond the predictive scope of conventional analytical models.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Physics-Informed Neural Networks, konduksi panas steady-state, media berpori, Metal-Organic Framework, konduktivitas termal efektif, Physics-Informed Neural Networks, steady-state heat conduction, porous media, Metal-Organic Framework, effective thermal conductivity
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QC Physics > QC320 Heat transfer T Technology > TP Chemical technology > TP1140 Polymers
Divisions:	Faculty of Industrial Technology > Material & Metallurgical Engineering > 28201-(S1) Undergraduate Thesis
Depositing User:	Rayhan Achmad Garibaldi
Date Deposited:	24 Jul 2026 06:22
Last Modified:	24 Jul 2026 06:22
URI:	http://repository.its.ac.id/id/eprint/137146

Actions (login required)

View Item