Neural networks as a tool for compact representation of ab initio molecular potential energy surfaces

Ab initio quantum chemical calculations of molecular properties such as, e.g., torsional potential energies, require massive computational effort even for moderately sized molecules, if basis sets with a reasonable quality are employed. Using ab initio data on conformational properties of the cofactor (6R,1'R,2'S)-5,6,7,8-tetrahydrobiopterin, we demonstrate that error backpropagation networks can be established that efficiently approximate complicated functional relationships such as torsional potential energy surfaces of a flexible molecule. Our pilot simulations suggest that properly trained neural networks might provide an extremely compact storage medium for quantum chemically obtained information. Moreover, they are outstandingly comfortable tools when it comes to making use of the stored information. One possible application is demonstrated, namely, computation of relaxed torsional energy surfaces.