Generative deep learning model assisted multi-objective optimization for wastewater nitrogen to protein conversion by photosynthetic bacteria.
Journal:
Bioresource technology
Published Date:
May 19, 2025
Abstract
For decades, the photosynthetic bacteria (PSB)-based nitrogen treatment and valorization from wastewater have been explored. However, balancing nitrogen removal performance and resource recovery potential in PSB has remained a key unresolved issue for a long time. This study employed generative deep learning algorithms to achieve high-quality data generation, supporting multi-objective optimization in nitrogen removal, protein concentration, and nitrogen-to-protein conversion. In this study, the Variational Auto-Encoders model generated 5000 samples related to PSB nitrogen recovery, significantly enhancing the original dataset. The Elastic Neural Network (ENN) model showed better fitting results with the generated data. In single-objective evaluations, SHapley Additive exPlanations analysis identified the most important factors: carbon source, nitrogen source, and light type for total nitrogen (TN) removal; nitrogen source, nitrogen loading rate (NLR), and light type for protein concentration; nitrogen source, light type, and chemical oxygen demand (COD) for nitrogen conversion. Multi-objective optimization identified eight pareto front points, with the following input variable ranges: COD 3.42-7.48 g L, TN 0.22-0.37 g L, COD:TN ratio 9.28-33.22, hydraulic retention time 4.02-7.67 days, illuminance 967.71-1405.56 lx, and NLR 0.28-0.77 g L d. The pareto solutions were mostly achieved under Near Infrared (NIR) light. Validation experiments further supported these findings, showing that NIR light achieved nitrogen-to-protein conversion reaching 44 % of the removed nitrogen. Additionally, NIR light significantly enhanced gene expression related to ammonia assimilation and protein translation processes compared to white light. The proposed generative framework provided an innovative solution for multi-objective optimization of wastewater nitrogen valorization under limited data conditions.