Multimodal artificial intelligence (AI) for estimating the sinking velocity of microplastic-microalgae aggregates...

Area of Science:

• Marine Biology
• Environmental Science
• Artificial Intelligence

Background:

• Microplastics are widespread marine pollutants.
• Microplastics form aggregates with microalgae (MP-MA), affecting their vertical distribution.
• Sinking mechanisms of MP-MA aggregates are not fully understood.

Purpose of the Study:

• To develop a multimodal AI model for estimating MP-MA aggregate sinking velocity.
• To investigate the influence of microalgal traits and aggregate structure on sinking behavior.
• To provide an interpretable framework for understanding MP-MA aggregate dynamics.

Main Methods:

• A multimodal AI model integrating image, text, and biological data was developed.
• Pretrained BLIP encoders extracted features from microscopy images and species names.
• Explainable AI techniques (attention rollout, SHAP) were used for model interpretation.

Main Results:

• The AI model achieved high accuracy (R²=0.857 train, R²=0.567 validation), outperforming Stokes-based models (R²=0.11).
• Attention maps highlighted key aggregate regions influencing sinking.
• SHAP analysis identified cell wall type, plastic particle count, and swimming mode as significant factors.

Conclusions:

• Multimodal AI offers a powerful, interpretable tool for predicting MP-MA aggregate sinking.
• Understanding these factors is crucial for assessing the environmental fate of microplastics and microalgae.
• This research advances the study of marine aggregate dynamics and pollutant transport.