Amplify scientific discovery with artificial intelligence

SCIENCE sciencemag.org T echnological innovations are penetrating all areas of science, making predominantly human activities a principal bottleneck in scientific progress while also making scientific advancement more subject to error and harder to reproduce. This is an area where a new generation of artificial intelligence (AI) systems can radically transform the practice of scientific discovery. Such systems are showing an increasing ability to automate scientific data analysis and discovery processes, can search systematically and correctly through hypothesis spaces to ensure best results, can autonomously discover complex patterns in data, and can reliably apply small-scale scientific processes consistently and transparently so that they can be easily reproduced. We discuss these advances and the steps that could help promote their development and deployment. Applying AI to the practice of science is not new. AI pioneer and Nobel laureate Herbert Simon hypothesized that cognitive mechanisms involved in scientific discovery are a special case of general human capabilities for problem-solving and, with colleagues, developed systems in the 1970s and 1980s that demonstrated reasoning capabilities for analyzing scientific data ( 1). Also in the 1970s, Joshua Lederberg (another Nobel winner) and colleagues developed the DENDRAL system for analyzing mass spectrometry data in order to hypothesize molecular structures ( 2). More recent breakthroughs, such as robot scientists and software that formulates laws for complex dynamical systems, demonstrate broader applicability of AI techniques for scientific discovery ( 3). Over the past two decades, AI has seen accelerating scientific advances and concomitant commercial-sector successes because of advances on three fronts: steady scholarly advances, especially as success has increased the numbers of interested participants; Moore’s law and steady exponential increases in computing power; and exponential increases in, and broad availability of, relevant data in volumes never previously seen. Those scientific efforts that have leveraged AI advances have largely harnessed sophisticated machine-learning techniques to create correlative predictions from large sets of “big data.” Such work aligns well with the current needs of petaand exascale science. However, AI has far broader capacity to ac-