Deep learning for groundwater level simulation in unconfined aquifers across the contiguous United States: Analyzing simulations at multiple lead times and integrating groundwater signatures

Simulating groundwater level accurately ahead of time is critical for various practical reasons. In this study, we simulate daily groundwater levels using meteorological data at 1-day to 7-day lead times across 249 unconfined wells in the contiguous United States. Our objectives include: 1) comparing LSTM-based Seq2Seq and Seq2One models for our simulation tasks, 2) investigating the efficacy of addressing failed models (where Nash Sutcliffe-Efficiency (NSE) < 0) by using alternative training-validation splits, and 3) integrating groundwater signatures for a more thorough model evaluation process. Our results demonstrate satisfactory to good performance with our best performing models achieving median NSE scores of 0.744 and 0.603 for 1-day and 7-day lead simulations, respectively. Notably, we show that LSTM-based Seq2One model outperforms the Seq2Seq model at 1-day lead simulation, however the difference in their performance is not statistically significant for 4-day and 7-day lead simulation scenarios. We found that 14% of our models exhibit subpar performance, which may be attributed to modeling complex underlying groundwater systems without key anthropogenic forcing variables. And we show that simply changing the training-validation splits is generally not sufficient to address these failed models. Our analysis with groundwater signatures, which are statistical aggregates of groundwater level hydrographs, reveals that most groundwater dynamics are well captured by the model, and notable correlations between model NSE and groundwater signatures are found. These findings position us towards better interpreting the capabilities and limitations of our groundwater models.