This 1953 paper by Nicholas Metropolis, Arianna Rosenbluth, Marshall Rosenbluth, Augusta Teller, and Edward Teller, produced at the Los Alamos Scientific Laboratory, set out to compute the equation of state for a system of many interacting molecules. Direct calculation was hopeless because the number of possible molecular configurations is astronomically large, so the authors devised a way to sample configurations on the MANIAC computer rather than enumerate them.
Their method, now known as the Metropolis algorithm, generates a sequence of configurations in which each new candidate is accepted or rejected according to a simple rule based on the change in energy. Lower-energy moves are always accepted; higher-energy moves are accepted with a probability that falls off exponentially. Over many steps this biased random walk visits configurations in proportion to their true physical probability, so averages over the walk approximate averages over the whole system.
Although the paper was written for statistical physics, the underlying idea turned out to be far more general. It is the first practical instance of what became Markov chain Monte Carlo, a family of methods for drawing samples from complicated probability distributions that cannot be sampled directly.
That generality is why the paper matters well beyond chemistry. Markov chain Monte Carlo is now a workhorse of Bayesian statistics and machine learning, used to fit models whose answers cannot be written down in closed form, and almost every modern variant traces back to the acceptance rule introduced here.