<p>Parameter inversion of arch dams is crucial for safety assessment, but it is usually carried out via finite element (FE) model updating or black-box optimization, which can be computationally expensive, difficult to balance physics- and data-driven objectives, and unstable for multi-parameter thermo-mechanical inversion while underusing short-term quasi-static responses. This study proposes a unified multi-objective inversion framework based on physics-informed neural networks for thermo-mechanical parameter identification in arch dams. Differentiable thermal and mechanical governing equations, boundary conditions and observation misfits are embedded into an end-to-end architecture to enable synchronous inversion of parameters under explicit physical constraints. To better exploit both long-term and short-term components of the monitored responses, a sinusoidal-activation network with prior-informed multi-channel inputs (spatial location, zoning information and dam-related thermal features) is adopted, enhancing physical consistency and high-frequency response representation. An Actor–Critic deep reinforcement learning scheme provides adaptive loss weighting, while the non-dominated sorting genetic algorithm II refines Pareto-optimal solutions. A case study on a concrete arch dam shows that the framework significantly reduces overall and high-frequency errors compared with a traditional separation inversion method. FE verification yields a maximum temperature error of 0.167 °C and a maximum radial displacement error of 0.143 mm, with relative errors below 4.89%.</p>