11:00 AM - 11:15 AM
[SCG55-02] Examining the effectiveness of a physics-informed neural network for the inverse modeling of hydrothermal systems
Keywords:Physics-informed neural network, Inverse modeling, Hydrothermal system
Predicting the temperature-pressure conditions and fluid flow patterns at depth is vital for a better understanding of hydrothermal systems. Numerical simulations have been generally conducted for this purpose, however the appropriate parameters must be set in advance and automatic calibration may require an enormous amount of calculation. Deep learning techniques have recently been applied and shown its effectiveness for predicting temperature at depths of hydrothermal systems. Although this deep learning approach enables predictions with fewer assumptions compared with the predictions by the numerical simulation, the lack of physical plausibility of the predicted quantities has been a major drawback.
Recently, a framework called physics-informed neural network (PINN) has been proposed and gained attention. As PINN considers physics laws described as partial differential equations and boundary conditions in the loss function, the physical plausibility of the predicted quantities can be enhanced. PINN has been applied to many research fields including fluid dynamics, computational mechanics and material sciences. Further, PINN has been shown to be useful for the problems in solid earth science, such as modeling crustal deformation, seismic tomography, and groundwater flow systems.
In this study, I proposed a PINN for predicting temperature, pressure and permeability distribution of a hydrothermal system, and examined its effectiveness with synthetic data. Specifically, I assumed sparse observations of temperatures, pressures and permeabilities at boreholes, and used the proposed PINN to predict the spatial distribution of the observed quantities. The accuracy and physical validity of the predicted quantities returned by the proposed PINN were compared with those predicted by a standard neural network without considering physics laws. For synthetic data, I simulated the natural-state temperature and pressure pattern from the permeability structure based on the geological structure of the Lahendong geothermal field, Indonesia.
The results showed that the predictions by PINN better satisfied the mass and energy conservation laws than the predictions by NN, thus showing superior physical validity. Moreover, the predicted temperature error at depth decreased by using the PINN compared with those returned by the NN. In addition, I found that the boundary condition settings played a fundamental role to decrease the prediction error, and the borehole locations were an important factor for the accurate prediction of the geological boundaries. The results in this study demonstrated the effectiveness of the proposed PINN for the inverse modeling of hydrothermal systems.
The author acknowledges the supports by JST ACT-X “AI powered research innovation” (grant no. JPMJAX20A1).
Recently, a framework called physics-informed neural network (PINN) has been proposed and gained attention. As PINN considers physics laws described as partial differential equations and boundary conditions in the loss function, the physical plausibility of the predicted quantities can be enhanced. PINN has been applied to many research fields including fluid dynamics, computational mechanics and material sciences. Further, PINN has been shown to be useful for the problems in solid earth science, such as modeling crustal deformation, seismic tomography, and groundwater flow systems.
In this study, I proposed a PINN for predicting temperature, pressure and permeability distribution of a hydrothermal system, and examined its effectiveness with synthetic data. Specifically, I assumed sparse observations of temperatures, pressures and permeabilities at boreholes, and used the proposed PINN to predict the spatial distribution of the observed quantities. The accuracy and physical validity of the predicted quantities returned by the proposed PINN were compared with those predicted by a standard neural network without considering physics laws. For synthetic data, I simulated the natural-state temperature and pressure pattern from the permeability structure based on the geological structure of the Lahendong geothermal field, Indonesia.
The results showed that the predictions by PINN better satisfied the mass and energy conservation laws than the predictions by NN, thus showing superior physical validity. Moreover, the predicted temperature error at depth decreased by using the PINN compared with those returned by the NN. In addition, I found that the boundary condition settings played a fundamental role to decrease the prediction error, and the borehole locations were an important factor for the accurate prediction of the geological boundaries. The results in this study demonstrated the effectiveness of the proposed PINN for the inverse modeling of hydrothermal systems.
The author acknowledges the supports by JST ACT-X “AI powered research innovation” (grant no. JPMJAX20A1).