| 電子情報通信学会 研究会発表申込システム 講演論文 詳細 |
技報閲覧サービス [ログイン] 技報アーカイブ |
| トップに戻る | 前のページに戻る | [Japanese] / [English] |
| 講演抄録／キーワード | ||
| 講演名 | 2025-09-17 16:15 ［依頼講演］Ray Tracing-Based Physics-Informed Machine Learning for Pathloss Prediction ○Inocent Calist・Minseok Kim（Niigata Univ.）  AP2025-82 |
|
| 抄録 | （和） | Accurate path loss (PL) prediction is essential for designing and optimizing next-generation wireless communication systems, particularly in complex urban environments where diverse propagation phenomena significantly impact signal behavior. This paper introduces a novel Ray Tracing (RT)-based Physics-Informed Machine Learning (PIML) framework for high-fidelity prediction of PL, utilizing a two-stage U-Net architecture. The framework synergistically combines RT simulation data, which captures fundamental electromagnetic propagation physics according to geometrical optics (GO) principles, with empirical PL measurements to integrate both physics-driven and data-driven insights. The pipeline encodes environmental features, including building density maps, log-distance fields, transmitter (Tx) and receiver (Rx) positions as multi-channel input images for the U-Net. The first stage learns physics-based propagation patterns from RT simulations, while the second stage refines predictions via residual learning using measured data. All inputs and outputs are spatially aligned on a 2048 × 2048 grid, and the network is optimized using a composite loss function that incorporates mean squared error and physics-informed constraints, ensuring adherence to wave propagation principles. Experimental evaluations conducted on real-world urban datasets from Yokohama, Japan, demonstrate the efficacy of the proposed ray-tracing (RT)-based physics-informed machine learning (PIML) framework in PL prediction. On validation data for route R3, the PIML model achieves a root mean square error (RMSE) of 4.42 dB and a coefficient of determination (R2) of 0.979, outperforming conventional RT-only simulations (RMSE = 6.58 dB, R2 = 0.953). While a purely data-driven U-Net model yields slightly better performance on R3 (RMSE = 3.55 dB, R2 = 0.986), the PIML framework exhibits superior generalization capabilities on an unseen route R4, with an RMSE of 18.34 dB compared to 18.52 dB for the data-driven model. This highlights the enhanced transferability of the hybrid approach. Furthermore, the modular architecture of the PIML framework enables efficient adaptation to diverse urban environments with minimal retraining. These findings underscore the potential of physics-informed neural architectures for robust, interpretable, and generalizable radio environment mapping in next-generation wireless communication systems. |
| （英） | Accurate path loss (PL) prediction is essential for designing and optimizing next-generation wireless communication systems, particularly in complex urban environments where diverse propagation phenomena significantly impact signal behavior. This paper introduces a novel Ray Tracing (RT)-based Physics-Informed Machine Learning (PIML) framework for high-fidelity prediction of PL, utilizing a two-stage U-Net architecture. The framework synergistically combines RT simulation data, which captures fundamental electromagnetic propagation physics according to geometrical optics (GO) principles, with empirical PL measurements to integrate both physics-driven and data-driven insights. The pipeline encodes environmental features, including building density maps, log-distance fields, transmitter (Tx) and receiver (Rx) positions as multi-channel input images for the U-Net. The first stage learns physics-based propagation patterns from RT simulations, while the second stage refines predictions via residual learning using measured data. All inputs and outputs are spatially aligned on a 2048 × 2048 grid, and the network is optimized using a composite loss function that incorporates mean squared error and physics-informed constraints, ensuring adherence to wave propagation principles. Experimental evaluations conducted on real-world urban datasets from Yokohama, Japan, demonstrate the efficacy of the proposed ray-tracing (RT)-based physics-informed machine learning (PIML) framework in PL prediction. On validation data for route R3, the PIML model achieves a root mean square error (RMSE) of 4.42 dB and a coefficient of determination (R2) of 0.979, outperforming conventional RT-only simulations (RMSE = 6.58 dB, R2 = 0.953). While a purely data-driven U-Net model yields slightly better performance on R3 (RMSE = 3.55 dB, R2 = 0.986), the PIML framework exhibits superior generalization capabilities on an unseen route R4, with an RMSE of 18.34 dB compared to 18.52 dB for the data-driven model. This highlights the enhanced transferability of the hybrid approach. Furthermore, the modular architecture of the PIML framework enables efficient adaptation to diverse urban environments with minimal retraining. These findings underscore the potential of physics-informed neural architectures for robust, interpretable, and generalizable radio environment mapping in next-generation wireless communication systems. | |
| キーワード | （和） | Path Loss / Physics-Informed Neural Networks / OpenStreetMap / Ray Tracing / U-Net / / / |
| （英） | Path Loss / Physics-Informed Neural Networks / OpenStreetMap / Ray Tracing / U-Net / / / | |
| 文献情報 | 信学技報, vol. 125, no. 180, AP2025-82, pp. 55-60, 2025年9月. | |
| 資料番号 | AP2025-82 | |
| 発行日 | 2025-09-10 (AP) | |
| ISSN | Online edition: ISSN 2432-6380 | |
| 著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します．(許諾番号：10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) | |
| PDFダウンロード | AP2025-82 |
|
| 研究会情報 | |
| 研究会 | MW AP |
| 開催期間 | 2025-09-16 - 2025-09-19 |
| 開催地（和） | 東京農工大学 小金井キャンパス |
| 開催地（英） | Koganei Campus, Tokyo University of Agriculture and Technology |
| テーマ（和） | 学生特集, ミリ波/マイクロ波, 電波伝搬, ⼀般 |
| テーマ（英） | Student Session, Microwave, Millimeter wave, Radio propagation, Antennas and Propagation |
| 講演論文情報の詳細 | |
| 申込み研究会 | AP |
| 会議コード | 2025-09-MW-AP |
| 本文の言語 | 英語 |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | Ray Tracing-Based Physics-Informed Machine Learning for Pathloss Prediction |
| サブタイトル（英） | |
| キーワード(1)（和/英） | Path Loss / Path Loss |
| キーワード(2)（和/英） | Physics-Informed Neural Networks / Physics-Informed Neural Networks |
| キーワード(3)（和/英） | OpenStreetMap / OpenStreetMap |
| キーワード(4)（和/英） | Ray Tracing / Ray Tracing |
| キーワード(5)（和/英） | U-Net / U-Net |
| キーワード(6)（和/英） | / |
| キーワード(7)（和/英） | / |
| キーワード(8)（和/英） | / |
| 第1著者 氏名（和/英/ヨミ） | Inocent Calist / Inocent Calist / |
| 第1著者 所属（和/英） | Niigata University (略称： Niigata Univ.) Niigata University (略称： Niigata Univ.) |
| 第2著者 氏名（和/英/ヨミ） | Minseok Kim / Minseok Kim / |
| 第2著者 所属（和/英） | Niigata University (略称： Niigata Univ.) Niigata University (略称： Niigata Univ.) |
| 講演者 | 第1著者 |
| 発表日時 | 2025-09-17 16:15:00 |
| 発表時間 | 25分 |
| 申込先研究会 | AP |
| 資料番号 | AP2025-82 |
| 巻番号（vol） | vol.125 |
| 号番号（no） | no.180 |
| ページ範囲 | pp.55-60 |
| ページ数 | 6 |
| 発行日 | 2025-09-10 (AP) |
[研究会発表申込システムのトップページに戻る]