Researchers in the US have created a computer program that can detect small tremors and distinguish minor earthquakes from background noise. Coordinated by seismologist Marine Denolle from Harvard University, USA, computer scientist Thibaut Perol and mathematician Michaël Gharbi used a type of artificial intelligence known as deep learning to create the software, which is capable of quickly analyzing large volumes of ground motion data called seismograms. The program, named ConvNetQuake, studies historic data examples and learns to recognize the unique characteristics of seismic waves produced by earthquakes, as well as differentiating them from the background noise generated by other phenomena, such as a truck passing near the detector. There are other existing programs that use machine learning to identify earthquakes, but unlike ConvNetQuake, they analyze the entire seismic wave—not just the relevant points. This requires more computer processing time, reducing the number of pattern comparisons and thus decreasing the number of earthquakes detected. The researchers tested the new system using data from seismic stations in Guthrie, Oklahoma, and detected 17 times more earthquakes than recorded in the state’s geological survey. In July 2014 alone, ConvNetQuake identified 4,225 earthquakes, most of them of low magnitude, in addition to those cataloged by the geological survey (Science Advances, February 14). In comparison with two other programs that use artificial intelligence, ConvNetQuake took 1 minute to analyze a week’s data—the second program took 48 minutes, and the third took nine days. Of the three, ConvNetQuake was the only one able to locate the origin of the tremor, with an accuracy of 74%.