We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Multimodal artificial intelligence (AI) for estimating the sinking velocity of microplastic–microalgae aggregates using explainable AI and visualization
Original title: Multimodal artificial intelligence (AI) for estimating the sinking velocity of microplastic–microalgae aggregates using explainable AI and visualization
Summary
When tiny plastic particles clump together with algae in the ocean, this affects whether they sink to the seafloor or float around where fish (and eventually humans) can encounter them. Researchers built an AI tool that predicts how fast these plastic-algae clumps sink, using photos, algae species info, and traits like cell wall type and swimming ability—and it worked far better than older math-based methods. This matters because understanding where microplastics end up in the ocean helps scientists track how they move through the food chain and eventually reach the seafood we eat.
Microplastics are pervasive pollutants in marine environments that interact with microalgae to form microplastic-microalgae (MP-MA) aggregates. These aggregates can increase effective density, thereby altering the vertical distribution and environmental fate of both microplastics and microalgae. However, the sinking mechanisms of these aggregates are not fully understood, partly due to their irregular morphology and species-specific traits. In this study, a multimodal artificial intelligence (AI) model was developed to estimate the sinking velocity of MP-MA aggregates using a dataset comprising 23 microalgal species. The model incorporated three input modalities: image data (microscopy images), text data (species names), and biological data (species-specific traits and experimental variables). Pretrained Bootstrapping Language-Image Pre-training (BLIP) encoders were used to extract image and text features, which were then concatenated with biological data for regression. Explainable AI and visualization techniques were applied to interpret the model results, including attention rollout visualization and SHapley Additive exPlanations (SHAP). The multimodal AI model outperformed both the Stokes-based formulations and the random forest model, achieving R values of 0.857 ± 0.051 for training and 0.567 ± 0.062 for validation across 20 different within-species training-validation data splits, whereas the best-performing Stokes-based formulation achieved an R of 0.11. Attention rollout visualization provided qualitative maps highlighting salient regions of the aggregates in the microscopy images. SHAP analysis identified cell wall type, plastic particle count, and swimming mode as influential variables. Overall, this study provides an interpretable basis for understanding how microalgal traits and aggregate structure relate to the sinking behavior of MP-MA aggregates in marine environments.