Abstract: As gas turbines progress towards higher efficiency and cleaner performance, the demand for combustion chamber designs grounded in high-fidelity models is becoming increasingly pressing. To meet this demand, a physics-informed neural network (PINN) based on variational auto-encoder flame (VAEF) was proposed for modeling both the forward and inverse problems of laminar premixed flames containing low-concentration chemical species. Following strategies were employed in this method: constrained mapping was implemented based on adaptive upper and lower boundaries or observational data; the computational domain was partitioned using smoothing windows with regional weighting; for inverse problems under sparse observation conditions, the variational auto-encoder reconstruction error and the Kullback-Leibler (KL) divergence term were incorporated into the calculation of the observation condition loss. Results indicate that these measures substantially enhance training stability and prediction accuracy for low-concentration species. A coefficient of determination greater than 0.95 and an relative error below 15% are achieved by PINN model, demonstrating excellent global fitting precision and parameter inversion capability. A PINN solution with better robust performance is thus provided for combustion simulation and emission monitoring under complex chemical mechanisms.