Machine learning and related computational methods have become substantially more powerful and are already applied in many areas of science. In the future, they are likely to change scientific research profoundly. In this talk I will be discussing two ways in which machine learning can be helpful in physics: solving differential equations and model building. I will attempt to explain the basic ideas behind these applications and present some recent examples, including inflationary model building, finding string models with certain prescribed properties and computing the masses of fermions from string theory.

We consider the heterotic string on certain six-dimensional half-flat manifolds which arise in the context of mirror symmetry. The four-dimensional N=1 effective theory is derived including the superpotential induced by the intrinsic torsion and NS NS flux. From the calculation of the gravitino mass term, we also obtain a general Gukov type formula valid for all heterotic compactifications on manifolds with SU(3) structure. The results are used to study heterotic AdS4 vacua which arise from the flux superpotential combined with gaugino condensation.