AI meets chemistry: new model predicts molecular Hessian matrices directly, skipping costly derivatives, 10-70 times faster

A University of Toronto team has published a new method, HIP, in Nature Computational Science that directly predicts a molecule's full Hessian matrix, skipping the costly differentiation step. Direct prediction: a single forward pass outputs the full Hessian matrix (3N×3N, second derivatives), with no automatic differentiation, finite differences or coupled-perturbed solves. Faster and cheaper: on small molecules it is 10-70 times faster than derivative-based machine-learning methods, uses 2-3 times less peak memory, and roughly halves the prediction error. Physics built into the architecture: based on the equivariant graph neural network EquiformerV2, rotational symmetry and matrix symmetry are guaranteed by construction rather than learned the hard way from training data. The Hessian is key to calculating molecular vibrations and reaction pathways; the faster it can be computed, the faster drug and materials screening can run. > Source: [read the original](https://bioengineer.org/ai-model-predicts-molecular-curvature-directly-skipping-costly-derivatives/). This text is a summary, for reference only.