We propose to explore an unprecedented amount of molecular data (166B molecules) with quantum chemical precision. For this, chemically accurate and transferable machine learning property models of unprecedented computational efficiency will be developed using special purpose tailored training sets designed according to unified multilevel techniques (sparse grids plus combination rules). We will subsequently apply the method to iteratively optimize molecular property objective functions which enable the routine discovery of new molecules with pre-defined specific properties – in real time. The objective for this effort is twofold: We plan to (a) provide experimental chemists with a powerful computational tool to guide design, synthesis, and characterization efforts of new and interesting molecules, and (b) gain a better understanding of the nature, landscape, and relationships among chemical structure and properties throughout chemical space.