Abstract: Accurate determination of temperature field in concrete gravity dams is essential for long-term operational safety. Most current solution methods suffer from mesh dependency issues. Although physics-informed neural networks (PINN) can directly incorporate physical laws into neural network training to achieve mesh-free solutions, they exhibit limitations in handling multi-medium interface problems. This study employed a discontinuity-capturing physics-informed neural network (DCPINN) model to solve the temperature field problem in heterogeneous roller-com pacted concrete dams. By introducing an additional dimension as the material identifier coordinate, the model effectively handled physical continuity at interfaces and overcame the limitations of traditional PINN in handling multi-medium interface problems. It also avoided the need to construct and train separate sub-networks for each sub-region required in domain-decomposition-type PINN, and eliminated the complexity and coordination challenges associated with multi-network joint optimization, reducing the overall model complexity. The applicability of DCPINN to solving temperature field problems in heterogeneous roller-compacted concrete dams was investigated, and the influence of different activation functions and network architectures on computational results was discussed. Numerical results show that the Tanh activation function is better suited to the DCPINN framework and that the neural network structure affects the final solution accuracy. DCPINN demonstrates high computational accuracy and reliability for addressing temperature-field problems in heterogeneous roller-compacted concrete dams and shows potential for engineering applications.