Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning

Abstract

Computational modeling of chemical and biological systems at atomic resolution is a crucial tool in the chemist’s toolset. The use of computer simulations requires a balance between cost and accuracy: quantum-mechanical methods provide high accuracy but are computationally expensive and scale poorly to large systems, while classical force fields are cheap and scalable, but lack transferability to new systems. Machine learning can be used to achieve the best of both approaches. Here we train a general-purpose neural network potential (ANI-1ccx) that approaches CCSD(T)/CBS accuracy on benchmarks for reaction thermochemistry, isomerization, and drug-like molecular torsions. This is achieved by training a network to DFT data then using transfer learning techniques to retrain on a dataset of gold standard QM calculations (CCSD(T)/CBS) that optimally spans chemical space. The resulting potential is broadly applicable to materials science, biology, and chemistry, and billions of times faster than CCSD(T)/CBS calculations.

Pure non-local machine-learned density functional theory for electron correlation

“…The main advantage of this direct approach is that electronic structure calculations are only required for training, whereas predictions can be performed using just the atomic coordinates as input. On the other hand, this also means that the complex physics underlying the PES has to be learned completely from data, which often translates to very large training sets from tens of thousands to millions of configurations, depending on the system (see Supplementary Note 5 ) 67 , 68 . Another downside is that most such ML forcefields are by construction short-ranged, meaning that long-range Coulomb and dispersion interactions are not included.…”

Dataset’s chemical diversity limits the generalizability of machine learning predictions

“…It is indeed a hot topic. Just in 2019, an impressive amount of studies have been devoted to the application of ML for the prediction of molecular energetic characteristics [1–11].…”

“…In chemistry-oriented ML, especially for molecular energies, the employed methods encompass Kernel Ridge Regression (KRR) [12–18], sometimes Gaussian process regression (GPR), linear regressions (Elastic Net, Bayesian Ridge Regression) and Random Forest (RF) [17, 19, 20]. The most recent state-of-the-art predictions have been obtained by neural networks (NN) [1, 2, 21–28]. ML performance as well depends on the molecular representation used.…”

“…ANI has been shown to predict total molecular energies at the level of < 1.5 kcal/mol albeit with a much larger training set size (a subset of GDB-11). This model has been very recently refined and errors in total energies prediction equal to 0.14 kcal/mol were reported by Smith et al [1] Schütt et al have proposed a deep tensor neural network (DTNN) to mimic many-body Hamiltonians [34]. Then they have introduced continuous filter convolutional layers as novel building blocks for deep neural networks [24, 35].…”

AP-Net: An Atomic-Pairwise Neural Network for Smooth and Transferable Interaction Potentials

“…energy prediction). [6][7][8][9][10][11][12][13][14][15][16][17][18] These ML potentials possess one commonality, which is that their development and application have emphasized covalently bound systems and accurate total energy predictions. Less attention has been paid to noncovalent interactions (NCIs) and interaction energies, which are of fundamental importance to drug binding, liquid structure, biomolecular structure, molecular crystals, etc.…”