Machine learning prediction and exploration of phosphorus migration and transformation during hydrothermal treatment of biomass waste

Abstract

Hydrothermal treatment (HTT) held promise for phosphorus (P) recovery from high-moisture biomass. However, traditional experimental studies of P hydrothermal conversion were time-consuming and labor-intensive. Thus, based on biomass characteristics and HTT parameters, Random Forest (RF) and Gradient Boosting Regression machine learning (ML) models were constructed to predict HTT P migration between total P in hydrochar (TP_HC) and process water (TP_PW) and hydrochar P transformation among inorganic P (IP_HC), organic P (OP_HC), non-apatite inorganic P (NAIP_HC), and apatite P (AP_HC). Results demonstrated that the RF models (test R2 > 0.86) exhibited excellent performance in both single-target and multi-target predictions. Feature importance analysis identified TP_feed, O, C, and N as critical features influencing P distribution in hydrothermal products. TP_feed, NAIP_feed, temperature, and IP_feed were crucial factors affecting P form transformation in HC. This study provided valuable insights into understanding the migration and transformation of P and further guided experimental research.