Force field inverse problems using recurrent neural networks

A new procedure to solve the force field inverse problem, based on recurrent neural networks, is discussed in this work. The present approach is numerically stable, simple and was able to recover reliable force constants for an error up to 30% in the initial condition. Force constants for water and benzene molecules, together with their main isotopes, are obtained on symmetry coordinates basis, from experimental frequencies. Comparison of these results with experimental and previous theoretical calculation is also made.