The high-throughput screening of large databases of novel materials candidates
constitutes a central goal of the Materials Genome Initiative (MGI). In previous
work we have shown that given a training set of molecules with pre-calculated
quantum mechanical properties, supervised machine learning models can be
constructed that infer corresponding properties for other molecules with quantum
chemical accuracy (~10 meV error). When compared to quantum mechanical
calculations, these estimates can be made at a computational cost reduced by
several orders of magnitude. Given sufficient CPU time, this approach could be
used for multiple property optimization runs that visit thousands, if not millions
of compounds. We propose to develop this approach so that it can be applied to the
high-throughput screening of novel condensed systems. If successful, the resulting
impact on atomistic simulation methodology as well as computational materials
design can be hardly overstated.