Corn dust is a highly energetic substance and frequently found in the food manufacturing industries. It not only poses occupational safety hazards such as suffocation or lung disorders for exposed persons but is often extremely explosible in ignition sensitive environment. This probability of explosion can be assessed and minimised with in-depth knowledge of controlling parameters/physical properties that trigger the ignition. This research takes into account the minimum ignition temperature (MIT), which is the control parameter for explosion risk assessment. MIT relies on multiple factors, such as moisture content, particle size, dust concentration, dispersion pressure, humidity and environmental temperature. In this study, the ignition of corn dust clouds was analysed using a Godbert Greenwald furnace for different combinations of dispersion pressure and concentrations. Test findings revealed that the minimum ignition temperature rises with a decrease in particle size. However, the minimum ignition temperature decreases with increased dispersion pressure and concentration until a specific value known as optimal value for ignition. Moreover, this work focuses on a statistical approach of polynomial surface fitting to forecast the MIT based on the combined impact of concentration and dispersion pressure on MIT for corn dust in a real-time experiment. The minimum value of the Bayesian Information Criterion (BIC) was used to select the most appropriate polynomial model due to its authenticity and strong reputation. An artificial neural network (ANN) is also used as a predictive tool to develop a model that can forecast the MIT with a defined combination of dispersion pressures and corn dust concentrations. As soon as an appropriate estimation of this minimum ignition temperature of the combustible dust is confirmed, it is possible to ensure that the temperatures of the surrounding hot surfaces do not rise to that point to prevent the explosion. The predictive results obtained from ANN were found to be good when compared with the polynomial surface fit. Several models with different numbers of neurons have been trained with different transfer functions. For the training, validation, and test phases, R values are around 1.0, i.e., 0.9863, 0.9930, and 0.9893, respectively. The overall R value was 0.9875 for the proposed network. The findings were considered to be acceptable as the overall value of R was close to 1.0. The network obtained sufficiently comparable findings with the research conducted by Kasalova and Balog. © 2021 Institution of Chemical Engineers