In order to accurately monitor CO2 concentration based on the non-dispersive infrared technique, a novel flat conical chamber CO2 gas sensor is proposed and investigated by simulation analysis and experimental verification. First, the optical design software and computational fluid dynamics method are utilized to theoretically investigate the relationship between the energy distribution, absorption efficiency of infrared radiation, and chamber size. The simulation results show that the chamber length has an optimal value of 8 cm when the cone angle is 5° and the diameter of the detection surface is 1 cm, which makes infrared absorption efficiency optimal. Then, the flat conical chamber CO2 gas sensor system is developed, calibrated, and tested. The experimental results indicate that the sensor can accurately detect CO2 gas concentrations in the range of 0–2000 ppm at 25 °C. It is found that the absolute error of calibration is within 10 ppm, and the maximum repeatability and stability errors are 5.5 and 3.5%, respectively. Finally, the genetic neural network algorithm is presented to compensate for the output concentration of the sensor to solve the problem of temperature drift. Experimental results demonstrate that the relative error of the compensated CO2 concentration is varied from −0.85 to 2.32%, which is significantly reduced. The study has reference significance for the structural optimization of the infrared CO2 gas sensor and the improvement of the measurement accuracy.