This research develops a physics-informed diffusion model for probabilistic solar irradiance forecasting that fundamentally integrates atmospheric thermodynamics into the generative modeling framework. Unlike conventional data-driven approaches that treat atmospheric processes as black boxes, our method explicitly leverages the Beer-Lambert law of atmospheric attenuation and thermodynamic principles to derive a physics-based noise schedule from atmospheric complexity functions. The model architecture employs a one-dimensional U-Net with cross-attention meteorological conditioning to predict atmospheric noise while enforcing physical constraints including energy conservation, thermodynamic consistency, and atmospheric bounds through specialized loss functions. The approach enables computationally efficient inference through DDIM sampling techniques and provides rigorous uncertainty quantification through multiple stochastic realizations. This work bridges the gap between purely data-driven machine learning techniques and physically-grounded atmospheric models, offering both predictive accuracy and scientific interpretability crucial for renewable energy grid integration and planning applications.