Kernel ridge regression (KRR) that satisfies energy conservation is a popular approach for predicting forcefield and molecular potential, to overcome the computational bottleneck of molecular dynamics simulation. However, the computational complexity of KRR increases cubically as the product of the number of atoms and simulated configurations in the training sample, due to the inversion of a large covariance matrix, which limits its applications to the simulation of small molecules. Here, we introduce the atomized force field (AFF) model that requires much less computational costs to achieve the quantum-chemical level of accuracy for predicting atomic forces and potential energies. Through a data-driven partition on the covariance kernel matrix of the force field and an induced input estimation approach on potential energies, we dramatically reduce the computational complexity of the machine learning algorithm and maintain high accuracy in predictions. The efficient machine learning algorithm extends the limits of its applications on larger molecules under the same computational budget. Using the MD17 dataset and another simulated dataset on larger molecules, we demonstrate that the accuracy of the AFF emulator ranges from 0.01-0.1 kcal mol$^{-1}$ or energies and 0.001-0.2 kcal mol$^{-1}$ $require{mediawiki-texvc}$$AA^{-1}$ for atomic forces. Most importantly, the accuracy was achieved by less than 5 minutes of computational time for training the AFF emulator and for making predictions on held-out molecular configurations. Furthermore, our approach contains uncertainty assessment of predictions of atomic forces and potentials, useful for developing a sequential design over the chemical input space, with nearly no increase of computational costs.
Download