Souvik and Miguel present at WEST conference 2025

Accurate forecasting of lake surface temperatures is essential for understanding freshwater ecosystem dynamics, managing water resources, and assessing regional climate impacts. Stannard Rock, a remote deep-water monitoring station in Lake Superior, provides high-resolution meteorological and lake temperature data, enabling detailed analysis of lake-atmosphere interactions. While Long Short-Term Memory (LSTM) neural networks have shown promise in capturing complex temporal patterns, they often lack physical interpretability and may struggle with sparse or incomplete data, especially during winter. In this study, we develop a physics-guided LSTM model tailored for forecasting daily lake surface temperatures at Stannard Rock. The model integrates local meteorological forcings—shortwave radiation, air temperature, wind speed, and longwave radiation—measured at the Stannard Rock buoy.

To promote physical consistency, we introduce a hybrid loss function that combines mean squared error with a penalty derived from the lake surface energy balance equation. This constraint penalizes deviations in heat flux dynamics, encouraging predictions that approximately conserve energy in accordance with thermodynamic principles. Temporal encodings are incorporated to capture seasonal variability. Initial training and validation are performed using a direct forecasting strategy, with seasonal data splits planned for future analyses. Model performance will be assessed using Root Mean Square Error (RMSE) and other time series accuracy metrics.

This methodology contributes to the growing field of physics-guided machine learning by embedding domain knowledge into data-driven models, offering a pathway to more interpretable and physically consistent forecasts of inland water dynamics.

 

Microbial Fuel Cells (MFCs) offer a sustainable solution to two global challenges: wastewater treatment and clean energy generation, by harnessing the metabolic activity of microorganisms to oxidize organic matter and produce electricity. However, their implementation remains limited due to system variability, power efficiency and scale-up challenges. Efficient modeling strategies are therefore essential to improve their performance and enable real-world applications.

In this study, we propose a hybrid modeling approach that combines a mechanistic one-dimensional model based on mass balances with machine learning (ML) techniques to simulate and predict the performance of a single-chamber MFC operating in batch mode. After calibrating the mechanistic model using experimental data, residual errors are used to train ML models in a residual learning framework. This approach enhances predictive accuracy while maintaining model interpretability. An experimental stage is included to generate the data required for model calibration and validation. Input variables include initial chemical oxygen demand (COD), pH, temperature, and time, while output variables include COD removal and cell voltage. The system is monitored over time to capture relevant process dynamics. This hybrid strategy supports the development of intelligent, adaptable MFC systems tailored to specific wastewater characteristics and operational conditions.

By integrating data-driven and physics-based methods, this work contributes to the advancement of low-energy treatment technologies aligned with circular economy principles and net-zero carbon objectives.