Breakthrough Digital Twin Model Predicts Arctic Permafrost Stability Under Roads

Permafrost beneath Arctic roads is warming and becoming less stable, posing increasing risks to northern infrastructure. Predicting how frozen ground will evolve is challenging due to sharply varying subsurface conditions, sparse observations, and the limitations of conventional process-based models.

In a new study published in the Journal of Geophysical Research: Earth Surface, Gou et al. [2026] introduce a physics-informed digital twin framework to address these challenges. The model was tested on an embankment road in Utqiaġvik, Alaska, using fiber-optic temperature measurements collected along a 100-meter transect to track shallow ground conditions over time.

Key Innovations in the Digital Twin Framework

The study’s framework integrates a neural network within a heat-transfer solver, ensuring that governing physics remain central while allowing the model to update uncertain soil properties as new data become available. This approach moves beyond traditional black-box prediction methods, enabling an interpretable and updateable system.

The digital twin can:

• Reconstruct subsurface temperature fields
• Infer thermodynamic properties such as unfrozen water content and thermal conductivity
• Test inferences against independent Distributed Acoustic Sensing (DAS) data, borehole temperatures, and laboratory measurements

These capabilities transform the model from a site-specific tool into a credible pathway for near-real-time permafrost forecasting and infrastructure monitoring in the rapidly warming Arctic.

How the Digital Twin Model Works

The framework’s structure is outlined in Figure 2 of the study. The neural network (NN) takes soil temperature at each lateral position as input and outputs six unknown parameters that vary laterally with distance. These parameters are embedded in the heat-transfer equation through constitutive relationships and solved using a finite difference method (FDM). The difference between predicted and observed temperatures is computed as “loss,” and loss gradients are backpropagated to update the NN parameters.

Visual Representation of the Framework

Figure 2: The neural network (NN) takes soil temperature at each lateral position as input and outputs six unknown parameters that vary laterally with distance. These parameters are embedded in the heat‐transfer equation through constitutive relationships, and the resulting system is solved using a finite difference method (FDM). The difference between predicted and observed temperatures is computed and defined as “loss,” and the loss gradients are backpropagated to update the NN parameters. Credit: Gou et al. [2026], Figure 2

Study Citation and Access

Gou, L., Xiao, M., Zhu, T., Martin, E. R., Wang, Z., Rocha dos Santos, G., et al. (2026). Physics-informed digital twin for predicting permafrost thermodynamic characteristics under an embankment road in Utqiaġvik, Alaska. Journal of Geophysical Research: Earth Surface, 131, e2025JF008787. https://doi.org/10.1029/2025JF008787

— Xiang Huang, Associate Editor, JGR: Earth Surface

Text © 2026. The authors. CC BY-NC-ND 3.0. Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.