Distribution coefficient prediction using multimodal machine learning based on soil adsorption factors, XRF, and XRD...

Area of Science:

• Environmental Science
• Geochemistry
• Nuclear Engineering

Background:

• The distribution coefficient (K_d) is critical for predicting radionuclide migration in soils.
• Existing models often fail to capture the complex interplay of geological and environmental factors influencing K_d, especially unique soil properties.
• Accurate K_d prediction is essential for environmental safety assessments around nuclear facilities.

Purpose of the Study:

• To develop a novel multimodal technique for predicting radionuclide distribution coefficients (K_d) in soils.
• To integrate diverse data sources, including physicochemical conditions, X-ray fluorescence (XRF), and X-ray diffraction (XRD) data.
• To provide a cost-effective and accurate method for assessing radionuclide adsorption mechanisms in soil environments.

Main Methods:

• Development and training of a multimodal model comprising three sub-networks.
• Sub-networks were designed to process distinct data domains: soil adsorption factors (physicochemical), XRF spectra, and XRD spectra (inherent soil properties).
• Model performance was evaluated using coefficient of determination (R²) and root mean squared error (RMSE) on natural log-transformed K_d values.

Main Results:

• The multimodal model demonstrated high predictive performance with R² = 0.84 and RMSE = 0.89 for log-transformed K_d.
• Identified influential peaks in XRD spectra related to inherent soil properties.
• Determined that soil pH and calcium oxide (CaO) content were significant variables from soil adsorption factors and XRF data, respectively.

Conclusions:

• This study presents the first multimodal model for simultaneously incorporating inherent soil properties and adsorption factors to predict K_d.
• The proposed technique offers a robust, cost-effective, and novel approach to understanding radionuclide adsorption in soils.
• Findings support the application of multimodal modeling for enhanced environmental safety and risk assessment at nuclear facilities.