A quantum-inspired attention integrated scalar long short-term memory model for accurate and stable groundwater...

In groundwater contamination source inversion, concentration data from monitoring wells serve as the most crucial known information, directly affecting the inversion accuracy of unknown contamination source parameters. However, existing studies often treat all monitoring data with equal importance, neglecting the varying contributions of individual wells to the inversion results. To address this issue, this study integrates the scaled dot-product attention (SDPA) mechanism into the scalar long short-term memory (sLSTM) model, forming a novel Attention-sLSTM inversion model. In this model, the attention module assigns different weights to monitoring wells, generating a weighted representation of the monitoring data. This weighted representation is then fed into the sLSTM model to learn the non-linear mapping between monitoring data and unknown source parameters. By highlighting critical monitoring data, the attention module improves the inversion accuracy of the sLSTM model. To further improve the attention module, two learnable parameters inspired by quantum theory are introduced into SDPA, yielding a quantum-inspired attention (QIA) mechanism. Compared to SDPA, QIA enables a more comprehensive representation of the complex relationships between monitoring data and contamination source parameters, leading to further gains in inversion accuracy. The proposed QIA-sLSTM model is evaluated through two synthetic groundwater contamination source inversion case studies. Results show that the QIA module consistently enhances inversion accuracy, while also showing the potential to improve model stability and reduce computational costs. Notably, the QIA-sLSTM model significantly outperforms the classical LSTM model, demonstrating superior and stable inversion performance across diverse groundwater contamination scenarios.