Heat and moisture simulations for buildings has been increasingly used in recent years to suggest appropriate retrofitting and operational methods. These simulations require various information such as material properties related to heat and moisture in the building structure, and flow characteristics related to ventilation and leakage. However, estimating these characteristics is not easy. The final goal of our study is to estimate these characteristics using a neural network trained by environmental measurement data in buildings. In this study, we proposed the neural network models estimating the ventilation rate in a single room as the first step. These models consist of two neural networks: one that estimates CO2 concentration changes in a room from time data and another that estimates ventilation rates using the estimated results of CO2 concentration. The former is optimized using a data-driven loss function, while the latter is optimized using a physics-informed loss function based on a mathematical model. In calculating the physics-informed loss, the time derivative of CO2 concentration is required, and this can be obtained through automatic differentiation by estimating CO2 concentration changes using a neural network. This method may ensure a certain degree of accuracy even when the measurement interval is long relative to the change rate in CO2 and the error due to numerical differentiation is large. If we apply this model to estimate ventilation rates in multi-room, it may be necessary to impose additional physical constraints that consider the flow balance throughout the building.