Key points are not available for this paper at this time.
Accurate reconstruction of subsurface temperature fields in layered media underpins exploration and development of deep geothermal resources. Traditional inverse computation methods improve numerical stability of finite‐difference schemes but still require careful regularization and layer‐by‐layer marching. In contrast, Physics‐Informed Neural Networks (PINNs) directly integrate governing equations, interface continuity, and boundary observations in a single mesh‐free optimization, dramatically reducing sensitivity to noise and eliminating the need for manual layer strategies. Through numerical experiments on two-dimensional multilayered domains, we show that PINNs method maintains robustness under realistic measurement noise, and deliver comparable accuracy without bespoke regularization parameters. Our results demonstrate that PINNs not only simplify the inverse workflow but also outperform classical layer‐marching approaches in accuracy, stability, and computational efficiency.
Liu et al. (Fri,) studied this question.