Traditionally, climate change prediction has been mainly based on physical calculations using supercomputers.However, the "bottom-up" conventional method of accumulating known physics knowledge, modeling and understanding the whole requires a finer simulation model and a larger supercomputer as more detailed calculations are made, which is a budget. There was a problem of expanding manpower.

　In contrast, this study incorporates "top-down" thinking that statistically analyzes trends.First, 30 consecutive years of temperature data are extracted from the past global temperature data, the temperature of each month from January to December is arranged vertically, and the temperature of each year of 1 years is arranged horizontally, and the temperature is colored by color. A pseudo-color image was generated.The reason for making images is that artificial intelligence makes it easier to learn its characteristics.

　By generating tens of thousands to hundreds of thousands of pseudo-color images and training them by deep learning, it is possible to predict whether the average temperature for the next 10 years will rise or fall with a maximum accuracy of 97% from past temperature data. It is said that it became.Actually, if you predict the average temperature for the next 2016 years from the temperature data up to 10, global warming will progress, but depending on the region, the temperature may rise slowly or may drop. The result of was obtained.Predicting regional differences in this way may contribute to the progress of climate research and global warming countermeasures in the future.

　In the future, this research group intends to further improve the accuracy of this method and integrate it with existing bottom-up climate forecasting.

Paper information:[Frontiers in Robotics and AI] Forecasting Climatic Trends Using Neural Networks: An Experimental Study Using Global Historical Data