Characterization of vapor intrusion sites with a deep learning-based data assimilation method

Area of Science:

• Environmental Science
• Geotechnical Engineering
• Data Science

Background:

• Accurate soil characterization is crucial for assessing soil vapor intrusion (VI) risks.
• Traditional methods often require extensive sampling, increasing costs and time.

Purpose of the Study:

• To develop and validate a deep learning-based data assimilation method (ES_(DL)) for estimating soil property distributions.
• To demonstrate the method's effectiveness in characterizing VI sites with limited data.

Main Methods:

• Development of the ES_(DL) data assimilation technique using deep learning.
• Application of the method to hypothetical VI scenarios.
• Validation through laboratory sandbox experiments and a real-world site case study.

Main Results:

• The ES_(DL) method provided reasonable estimations of effective diffusion coefficients and emission rates.
• Sufficient sampling (e.g., 15 points) effectively characterized spatial soil heterogeneity.
• Layered characterization is an alternative when horizontal heterogeneity cannot be captured.

Conclusions:

• The ES_(DL) method offers an efficient alternative for characterizing VI sites.
• It enables accurate risk assessment with reduced sampling requirements.
• The study highlights the potential of data assimilation and deep learning in environmental site assessment.