Air quality monitoring increasingly relies on low-cost sensors to densify the monitoring network of air pollutants and enhance the modelling of air pollution cost-effectively. However, these sensors require calibration due to their sensitivity to environmental conditions and inter-device variability. This study employs a random forest method to tackle these challenges by selecting environmental predictive variables and developing calibration models for sensors measuring particulate matter (PM2.5 and PM10). The random forest method effectively addresses the non-linear relationships between the concentration and five predictive variables (temperature, atmospheric pressure, relative humidity, wind speed, and direction). Three ministations, developed at ISSeP and equipped with Sensirion SPS30 sensors, were tested at INERIS under a hybrid-controlled atmosphere, using ambient air as the matrix while introducing controlled levels of gaseous and PM pollutants. Calibration models were developed using two datasets with varying pollutant concentrations. The results highlight the importance of incorporating meteorological variables to enhance model robustness, with the calibration period significantly influencing this robustness. Achieving high performance requires encompassing a representative range of predictive variable combinations in the calibration process. Various solutions have been proposed to address these challenges, underscoring the need for further investigation into the calibration of low-cost sensors.