The biological carbon pump (BCP) plays a crucial role in sequestering carbon from the atmosphere to the ocean. Sinking particulate organic carbon (POC) is the main component of BCP. However, the magnitude of POC flux calculated using different methods varies significantly. This study aims to narrow down this uncertainty by providing a new estimate using the disequilibrium between ²³⁴Th and ²³⁸U. Specifically, we train a Gaussian process regression model using a recent compilation of ²³⁴Th observations and eight environmental predictors including apparent oxygen utilization, dissolved oxygen (DO), density, nitrate, phosphate, salinity, silicate, temperature, which have a global distribution with monthly resolution. The estimate of global mean POC flux at the depth of 100 m is 5.36 ± 0.12 Pg C yr⁻¹, with strong seasonal variations. The Model suggests that DO, salinity, silicate, and temperature are the primary factors contributing to the seasonal and spatial variations of POC export. Additionally, by using the 2100 SSP5-8.5 data as inputs for the model, it is determined that POC flux will decrease by approximately 6%.
 

thorium-234, machine learning, particle dynamics