Quantifying the dynamics of peak innovation in scientific careers
We examine the innovation of researchers with long-lived careers in Computer Science and Physics. Despite the epistemological differences between such disciplines, we consistently find that a researcher's most innovative publication occurs earlier than expected if innovation were distributed at random across the sequence of publications in their career, and is accompanied by a peak year in which researchers publish other work which is more innovative than average. Through a series of linear models, we show that the innovation achieved by a researcher during their peak year is higher when it is preceded by a long period of low productivity. These findings are in stark contrast with the dynamics of academic impact, which researchers are incentivised to pursue through high productivity and incremental - less innovative - work by the currently prevalent paradigms of scientific evaluation.