We present a surrogate model for simulating post-exposure bake (PEB) in optical lithography. PEB is the thermal treatment that follows exposure of chemically amplified photoresists (CARs), during which acid diffusion and acid-catalyzed reactions convert the latent aerial image into a developable pattern. Accurate PEB simulation is therefore essential for semiconductor manufacturing, but typically accounts for 10-30% of the total runtime in lithography simulation workflows. PEB can be described mathematically by a coupled reaction-diffusion system for the concentrations of inhibitor, acid, and base. The corresponding model parameters depend on the specific resist chemistry and processing conditions (temperature, bake time, etc.) and are typically determined via calibration using repeated runs of a numerical solver. Here, we present an alternative approach based on a physics-informed neural operator (PINO). The PINO serves as a surrogate to model the modification of the photoresist composition during the PEB, taking the initial state as input and predicting the subsequent evolution of the concentrations. The PINO is implemented as a DeepONet in which different physical priors are directly encoded, including boundary conditions, high-frequency behaviour, and structural properties of the underlying equations. Training directly employs the governing physical laws through physics-informed loss terms, thereby removing the need for large amounts of simulation data, and is augmented by a small set of in-house measurement data. In this way, parameter calibration is embedded directly into the PINO's training, resulting in a surrogate model that is simultaneously constrained by physical laws and measurements. We demonstrate that the trained PINO has high accuracy in application-relevant lithography metrics, closely matching experimental results, while achieving speed-ups of multiple orders of magnitude over classical solvers.