Abstract: The melting point of slag plays a pivotal role in determining the process stability and product quality during electroslag remelting (ESR). However, accurate prediction remains challenging due to the scarcity of experimental data, the compositional complexity of slag systems, and the limited accuracy of conventional empirical models. In this study, a robust multi-task neural network framework based on multi-source data fusion is developed to predict the melting points of ESR slags. Three complementary data sources are integrated: (i) experimentally measured melting points, (ii) melting point estimations derived from viscosity inflection points using the viscosity prediction model, and (iii) synthetic data generated from empirical correlations. A multi-task learning architecture is constructed, comprising one primary task (experimental data) and two auxiliary tasks (augmented datasets), enabling effective knowledge transfer and joint optimization through a custom-designed loss function and adaptive task-weighting scheme. Experimental results demonstrate that the proposed method consistently outperforms traditional machine learning models and single-task neural networks across multiple evaluation metrics, including MSE, MAE, and MAPE. In addition, the model exhibits excellent robustness and physical consistency, providing a promising solution for intelligent slag system design and process optimization in ESR operations.