Prediksi Distribusi Temperatur Steady-State pada Material ZrP2O7 dan ZrP2O7-CePO4 Berpori 2D Menggunakan Physics-Informed Neural Networks (PINN)

Amarta, M. Adila (2026) Prediksi Distribusi Temperatur Steady-State pada Material ZrP2O7 dan ZrP2O7-CePO4 Berpori 2D Menggunakan Physics-Informed Neural Networks (PINN). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5011221177-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (17MB) | Request a copy

Abstract

Evaluasi perpindahan panas pada material berpori menggunakan metode numerik konvensional seperti Finite Element Method (FEM) sering kali terhambat oleh tingginya biaya komputasi akibat kebutuhan membangun ulang jaring spasial (remeshing) setiap kali topologi pori berubah. Penelitian ini mengusulkan penggunaan Physics-Informed Neural Networks (PINN) sebagai alternatif model komputasi berbasis machine learning untuk memprediksi distribusi temperatur keadaan tunak (steady-state) pada material isolator termal ZrP2O7 dan ZrP2O7-CePO4. Model dievaluasi pada domain dua dimensi dengan variasi tingkat porositas (20–60%), karakteristik ukuran pori (lebih kecil dan lebih besar), serta kerapatan titik kolokasi. Hasil simulasi menunjukkan bahwa peningkatan porositas secara konsisten memperparah fragmentasi jalur konduksi panas matriks padat sehingga meningkatkan resistensi termal dan menurunkan magnitudo fluks panas. Selain itu, pori berukuran kecil menghasilkan distribusi panas yang relatif merata, sedangkan pori berukuran besar bertindak sebagai rintangan termal makroskopis yang menciptakan fenomena bottlenecking, meningkatkan tortuositas termal, serta menyebabkan akumulasi panas lokal pada celah matriks yang sempit. Untuk mengatasi fenomena spectral bias pada area diskontinuitas batas pori, penelitian ini mengimplementasikan pendekatan transfer learning (TL) intra-geometri yang dikombinasikan dengan formulasi Hard Boundary Condition. Hasilnya menunjukkan bahwa implementasi transfer learning dari resolusi rendah menuju resolusi tinggi (≥150 titik kolokasi) mampu menekan Maximum Absolute Error (MaxAE) lebih dari 50% dibandingkan model konvensional dan menghasilkan kesesuaian yang sangat baik terhadap solusi FEM dengan Mean Absolute Percentage Error (MAPE) keseluruhan di bawah 1,1%. Dari sisi komputasi, meskipun fase pelatihan PINN membutuhkan waktu yang relatif besar hingga sekitar 1.700 detik, model ini mampu menghasilkan prediksi temperatur pada fase inferensi dalam waktu sekitar 1 milidetik tanpa memerlukan remeshing, sehingga sangat potensial untuk simulasi skala besar dan evaluasi berulang. Analisis lebih lanjut menunjukkan bahwa nilai konduktivitas termal efektif (keff) yang diperoleh dari prediksi PINN mengikuti tren teoritis Effective Medium Theory (EMT), yaitu menurun seiring peningkatan porositas akibat berkurangnya media padat yang berperan dalam proses konduksi panas.
=====================================================================================================================================
Heat transfer evaluation in porous materials using conventional numerical methods such as the Finite Element Method (FEM) is often hindered by high computational costs due to the need for remeshing whenever the pore topology changes. This study proposes the use of Physics-Informed Neural Networks (PINNs) as a machine learning-based computational framework for predicting steady-state temperature distributions in porous thermal insulating materials, namely ZrP2O7 and ZrP2O7-CePO4. The model was evaluated on two-dimensional domains with varying porosity levels (20–60%), pore size characteristics (small and large pores), and collocation point densities. The simulation results indicate that increasing porosity consistently intensifies the fragmentation of solid-phase heat conduction pathways, leading to higher thermal resistance and lower average heat flux. Furthermore, small pores produce relatively uniform heat distribution patterns, whereas large pores act as macroscopic thermal barriers that induce bottlenecking effects, increase thermal tortuosity, and generate localized heat accumulation within narrow matrix regions. To address the spectral bias phenomenon near pore boundary discontinuities, an intra-geometry transfer learning (TL) strategy combined with a Hard Boundary Condition formulation was implemented. The results demonstrate that transfer learning from low-resolution to high-resolution models (≥150 collocation points) reduces the Maximum Absolute Error (MaxAE) by more than 50% compared with conventional PINN models and achieves excellent agreement with FEM solutions, maintaining an overall Mean Absolute Percentage Error (MAPE) below 1.1%. From a computational perspective, although the PINN training phase requires substantial computational time, reaching approximately 1,700 seconds for the baseline model, the trained network performs thermal field inference in approximately 1 millisecond without remeshing, making it highly suitable for large-scale and repeated simulations. Furthermore, the effective thermal conductivity (keff) extracted from PINN-predicted temperature fields exhibits macroscopic trends consistent with Effective Medium Theory (EMT), showing a systematic decrease with increasing porosity due to the reduction of the solid medium available for heat conduction.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Konduktivitas Termal, Material Berpori, Perpindahan Panas, Physics-Informed Neural Network (PINN), Transfer Learning, Heat Transfer, Physics-Informed Neural Networks (PINNs), Porous Media,Thermal Conductivity, Transfer Learning
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QC Physics > QC320 Heat transfer T Technology > T Technology (General) > T57.62 Simulation
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Material & Metallurgical Engineering > 28201-(S1) Undergraduate Thesis
Depositing User:	M. Adila Amarta
Date Deposited:	24 Jul 2026 02:31
Last Modified:	24 Jul 2026 02:31
URI:	http://repository.its.ac.id/id/eprint/136690

Actions (login required)

View Item