Japan Geoscience Union Meeting 2022

Presentation information

[J] Poster

S (Solid Earth Sciences ) » S-CG Complex & General

[S-CG51] Driving Solid Earth Science through Machine Learning

Mon. May 30, 2022 11:00 AM - 1:00 PM Online Poster Zoom Room (27) (Ch.27)

convener:Hisahiko Kubo(National Research Institute for Earth Science and Disaster Resilience), convener:Yuki Kodera(Meteorological Research Institute, Japan Meteorological Agency), Makoto Naoi(Kyoto University), convener:Keisuke Yano(The Institute of Statistical Mathematics), Chairperson:Hisahiko Kubo(National Research Institute for Earth Science and Disaster Resilience), Yuki Kodera(Meteorological Research Institute, Japan Meteorological Agency), Keisuke Yano(The Institute of Statistical Mathematics)

11:00 AM - 1:00 PM

[SCG51-P06] Temperature estimation at depth by physics-informed neural network: accuracy evaluation with 2D numerical data

*Kazuya Ishitsuka1 (1.Kyoto University)

Keywords:Temperature estimation, Neural network considering a physical law, Thermal convection

Estimation of temperatures at depth is important for understanding physical processes in the Earth. Especially for geothermal applications, neural network approach has been used to estimate temperatures based on existing temperature logs. This approach trains the relationship between locations and temperatures at existing wells, and then predicts the temperatures at other locations. Although this approach has shown the effectiveness, the estimated temperatures do not necessarily follow physical laws. Recently, physics-informed neural network has been proposed, and evaluated the accuracy in groundwater applications. As this method constrains the predicted quantities using physical laws described by partial derivative functions, the predictions can follow the basic physical equations. In this study, I extended the proposed method for temperature estimation at depth and examined the prediction accuracy using 2D geothermal simulation data. For 2D simulation data, I considered a water saturated domain with the depth of 1 km and a constant heat flux at the bottom boundary. As a result, the RMSE of the estimated temperature was 7.7% using the physics-informed neural network, while the RMSE by a standard neural network was 25.2%. In addition, the loss of the mass conservation decreased by using the physics-informed neural network.

The author acknowledges the supports by JST ACT-X "AI powered research innovation" (grant no. JPMJAX20A1).