Advanced Deep Learning Models for Groundwater Contaminant Dynamics at the Savannah River Site

Advances in sensor technology have greatly enhanced the monitoring and assessment of environmental conditions. However, groundwater monitoring presents unique challenges, primarily in accurately detecting anomalies amidst sensor noise, fluctuations, and other inconsistencies in the data. The introduction of low-cost in-situ sensors has dramatically increased data volumes, requiring more advanced analytical techniques. This paper investigates the potential of Artificial Intelligence and Machine Learning (AI/ML), specifically Long Short-Term Memory (LSTM) Autoencoder models and Transformers, to tackle these challenges. LSTM Autoencoders, a specialized form of recurrent neural networks, are well-suited for sequential data analysis and anomaly detection, particularly in time-based sensor readings. The LSTM Autoencoder can compress the data and reconstruct normal patterns by training on historical sensor data representing normal groundwater conditions, thus identifying deviations that signal potential contamination. Meanwhile, transformer models, known for their attention mechanism and ability to capture long-range dependencies, are incorporated to enhance further their ability to distinguish between natural data variations and significant anomalies in complex, dynamic groundwater environments. We apply these LSTM Autoencoder and Transformer approaches to US-DoE Savannah River Site F-Area datasets, demonstrating their robustness and adaptability in identifying subtle and progressive groundwater contamination. This approach improves the accuracy of groundwater monitoring and supports a proactive stance in environmental preservation and public health protection.