Physics-informed multi-task learning framework for modelling microplastic impacts on sea turtle nest temperature and sex determination.

Microplastic (MP) pollution in coastal nesting beaches may alter the thermal environment of sea turtle nests and shift hatchling sex ratios through temperature-dependent sex determination (TSD). This study investigates how MP abundance and characteristics influence middle-third nest temperature and sex ratio in green turtles (Chelonia mydas) across three nesting sites in Terengganu, Malaysia. A physics-informed multi-task learning (PI-MTL) model was developed to jointly predict nest temperature and sex-ratio class, using detailed MP profiles (color, shape, and size) across different depth level distribution. By incorporating a physics-informed consistency loss derived from the pivotal temperature and transitional thermal bandwidths, the PI-MTL model outperformed established machine learning methods including conventional single-task learning (STL), k-nearest neighbors (kNN), and support vector machines (SVM), achieving the lowest total absolute temperature error (4.037 °C) and 100% sex-ratio classification accuracy. PI-MTL also generated more consistent predictions that remained coherent with TSD mechanism, whereas the other models occasionally yielded thermally and biologically inconsistent outputs. A SHapley Additive exPlanations (SHAP) feature importance analysis further revealed that color-related MP attributes, particularly black and blue-colored particles, exert the strongest influence on nest temperature predictions. This study demonstrates that integrating MP characterization with a PI-MTL framework, complemented by SHAP-based feature importance analysis, provides a robust and biologically grounded approach for predicting MP-driven shifts in nest temperature and sex ratio and informing targeted conservation strategies in data-limited contexts.