The process of glass formation—the transition from liquid to amorphous solid—has been known for centuries, and glasses have occupied an indispensable place in our lives ever since. Despite this abundance, however, the physics of glass formation remains notoriously poorly understood. Recently a promising new first-principles theory of the glass transition has been developed, but the corresponding equations are extremely computationally demanding to solve. In this project, a deep neural network has been trained to accurately reproduce the solutions of this theory within milliseconds. This orders-of-magnitude speedup suggests that machine-learning methods can provide a powerful alternative to complex numerical integration algorithms associated with complex physical theories.