Multi-task deep learning for pollutant source inversion with DFNN, LSTM, and Transformer architectures

Abstract

Gas leakage incidents in chemical industrial parks can lead to severe economic losses and pose significant risks to human safety. Rapid identification of the leakage source enables timely mitigation, while accurate estimation of the emission strength helps assess the severity of the incident. This study presents a multi-task learning (MTL) framework for source term estimation (STE) that simultaneously predicts source location and time-varying emission strength. Three representative deep learning architectures, a Deep Feedforward Neural Network (DFNN), a Long Short-Term Memory (LSTM) network, and a Transformer, are compared under both constant and dynamic release scenarios. This work provides the first systematic evaluation of these distinct architectures within an MTL framework for STE, demonstrating the advantages of temporal feature learning for inverse modeling applications. A realistic and large-scale dataset is generated using computational fluid dynamics (CFD) and the response factor method (RFM) to simulate dispersion. Optuna-based hyperparameter optimization is employed to ensure reliable model comparison. Results demonstrate that all three models achieve strong inversion performance. The DFNN proves efficient and robust in constant-release scenarios, while the LSTM excels under dynamic conditions, significantly improving the estimation accuracy over a shallow ANN without MTL, reducing the MAE for source strength from 0.394 to 0.147 and increasing the R${}^2$ from 0.284 to 0.768. Therefore, for time-varying emissions, the MTL-based LSTM is recommended due to its superior ability to capture temporal dynamics and provide precise rate estimates.