Simulating biology & materials science at the atomic scale using AI that inputs atomic coordinates and outputs the energy surface necessary answer relevant questions
Many methods in computational chemistry have been developed over the decades for simulating physical processing, especially in biology, but come with the tradeoff of speed vs accuracy. Fast methods that model only classical physics aren’t transferable to other systems and often cannot simulate complex phenomena (e.g. protein folding, covalent bond-breaking in drug binding or heterogenous battery chemistries in materials science) while the methods that directly calculate the breadth of inter-atomic effects necessary to predict such phenomena are excruciatingly slow. Neural net potentials embed only the crucial effects in their architecture explicitly and learn the remaining patterns implicitly from the training data so they never do the full calculations from scratch. They’re thus poised to dramatically expand the Pareto frontier, enabling highly accurate simulation of arbitrarily complex phenomena at speeds, system scale, and computational efficiency relevant for industrial use.
From this review paper
Key aspects of this system include:
The potential applications are limited only by the ultimate speed and accuracy breakthroughs enable.
This PhD dissertation of a leading researcher provides an excellent intro to the field, the theory its grounded in, and adjacent approaches. This talk provides a great overview of the NNPs' recent development more broadly. And, here are talks given about two of the top models, MACE and Allegro.