Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Back

4 Views

Pavements 2025

Machine Learning & AI Applications in Pavements

POSTER RECEPTION (View list of posters here)

Predicting Frost Penetration Depth using Recurrent Neural Networks and Worldwide Reanalysis Data

Tuesday, June 10, 2025

4:45 PM – 6:00 PM MT

Primary Author(s)

TS

Tavis Sotirin-Miller

Research Computer Scientist
Cold Regions Research and Engineering Laboratory, Engineer Research and Development Center, US Army Corps of Engineers
White River Junction, VT, United States

Presenter(s)

TS

Tavis Sotirin-Miller

Research Computer Scientist
Cold Regions Research and Engineering Laboratory, Engineer Research and Development Center, US Army Corps of Engineers
White River Junction, VT, United States

Co-Author(s)

Scott Michael M. Slone, n/a

Research Materials Engineer
USACE-ERDC
Hanover, New Hampshire, United States
RI

Robert Ibey, n/a

Research Physical Scientist
USACE-ERDC
Hanover, New Hampshire, United States
Wade A. Lein, PhD, PE

Research Civil Engineer
ERDC/CRREL
Enfield, New Hampshire, United States

Submitter(s)

TS

Tavis Sotirin-Miller

Research Computer Scientist
Cold Regions Research and Engineering Laboratory, Engineer Research and Development Center, US Army Corps of Engineers
White River Junction, VT, United States

Abstract: Ground frost can have serious effects on existing structures, foundations, and paved surfaces, as well as the design of future constructions. With the increase in granularity and availability of historical weather data through organizations such as the European Centre for Medium-Range Weather Forecasts (ECMWF), AI models are increasing in popularity and accuracy year over year, as weather data lends itself well to trend forecasting and training machine learning models. Utilizing global satellite data from the ECMWF ERA-5 database, a time-series recurrent neural network was developed and trained to predict frost penetration depth in various soil types. Inputs were local historical precipitation, temperature, and cloud coverage, as well as soil parameters, such as thermal conductivity, soil temperature at various depths, and soil texture. The model is able to predict future frost penetration depths and other soil parameters within a set time window, which could aid in existing infrastructure management as well as planning and development. The model is capable of operating in most cold regions in the world thanks to the global availability of the ERA-5 data and can be retrained to focus on specific soil types or areas if needed. A variety of types of neural network architectures, parameters, and organizations of data were tested and studied to fine-tune end model performance.

Learning Objectives:

Attendees can expect to learn the following from this session:

identify optimal neural network architectures for predicting ground frost penetration.
list relevant, globally available data for predicting ground frost penetration.
conduct custom neural network training for unique soil types on existing trained networks to better predict ground frost penetration in specific areas.