TECHNOLOGY

A digital twin for aquaculture: AI predicts fish growth in real time

China, 12 September 2025 | Traditional statistical models, and machine learning will be foundational components for next generation of forecasting tools based on physics-informed neural networks

Digital Twin en acuicultura

Traditionally, fish growth has been estimated using statistical models. More recently, machine learning has gained ground as a predictive tool. However, both approaches have limitations. Statistical models often struggle to adapt to the complexity and variability of aquaculture environments, while machine learning techniques act as “black boxes”, offering little explanation for their predictions.

The next generation of forecasting tools is based on physics-informed neural networks (PINNs). These models combine data-driven learning with biological principles to provide accurate and interpretable predictions. By embedding the laws of growth biology, they deliver forecasts that are not only precise but also biologically consistent—an essential requirement for aquaculture management.

A team of researchers from the Hubei University of Technology, China, has developed such a model, known as L-PIGRU (Long Short-Term Memory Physics-Informed Gated Recurrent Unit). By integrating neural networks with physics-informed constraints, the system ensures that predictions remain in line with established biological principles. According to the authors, this hybrid approach “enhances both the flexibility and rationality of the model”.

L-PIGRU combines two main components. The first is a Long Short-Term Memory Network (LSTM), which analyses historical weight data, feeding records and water quality indicators to identify temporal growth patterns. The second is a Physics-Informed Gated Recurrent Unit (PIGRU), which introduces constraints derived from the Dynamic Energy Budget (DEB) model, ensuring that predictions reflect key biological processes such as metabolism and feeding efficiency.

This dual design reduces dependence on very large datasets while improving both reliability and interpretability—two critical factors for decision-making in aquaculture.

The model was validated under controlled laboratory conditions in a recirculating aquaculture system (RAS) equipped with advanced monitoring sensors, where researchers reared largemouth bass (Micropterus salmoides). Eight independent datasets covering complete production cycles were used for training and testing.

The results were striking: L-PIGRU achieved a mean absolute percentage error (MAPE) of just 0.68%, compared with 4.01% for ARIMAX, 2.03% for LSTM, and 2.48% for PINN models.

These results establish L-PIGRU as the most accurate tool tested to date, while also maintaining transparency in its predictions.

Accurate and interpretable forecasts can provide direct support to farm management. With L-PIGRU, fish farmers could optimise feeding strategies, reduce waste, improve growth efficiency, and plan harvest schedules more effectively. The authors also highlight that this type of model lays the foundation for creating a digital twin of aquaculture operations, enabling real-time simulation and monitoring.

Related