Physics-informed neural network-based serial hybrid model capturing the hidden kinetics for sulfur-driven autotrophic...

Area of Science:

• Environmental Biotechnology
• Wastewater Treatment
• Bioreactor Modeling

Background:

• Sulfur-driven autotrophic denitrification (SdAD) removes nitrate from low C/N wastewater.
• Mechanisms of intermediate accumulation (elemental sulfur, nitrite) in SdAD are not fully understood.
• Existing models struggle to predict SdAD intermediates due to incomplete kinetic knowledge and environmental variability.

Purpose of the Study:

• To develop a novel hybrid model using physics-informed neural networks (PINNs) for SdAD.
• To accurately capture SdAD process kinetics and predict substrate concentrations.
• To address limitations in existing mathematical models for SdAD intermediate prediction.

Main Methods:

• Proposed a serial hybrid model structure integrating PINNs.
• Evaluated the model via numerical experiments.
• Applied the model to batch and continuous-flow reactor case studies.

Main Results:

• The PINN-based hybrid model achieved accurate state and kinetic predictions in numerical experiments.
• The model outperformed both mechanistic and purely data-driven models in real case studies.
• The trained model facilitated the design of control strategies for SdAD and integrated processes.

Conclusions:

• PINN-based hybrid models offer superior accuracy for predicting SdAD dynamics and intermediates.
• This approach enhances understanding and control of complex biological wastewater treatment processes.
• The model shows promise for designing energy-efficient nitrogen removal systems.