Transferable Neural Network Potential Energy Surfaces for Closed-Shell Organic Molecules: Extension to Ions

Abstract

Transferable high dimensional neural network potentials (HDNNPs) have shown great promise as an avenue to increase the accuracy and domain of applicability of existing atomistic force fields for organic systems relevant to life science. We have previously reported such a potential (Schrodinger-ANI) that has broad coverage of druglike molecules. We extend that work here to cover ionic and zwitterionic druglike molecules expected to be relevant to drug discovery research activities. We report a novel HDNNP architecture, which we call QRNN, that predicts atomic charges and uses these charges as descriptors in an energy model that delivers conformational energies within chemical accuracy when measured against the reference theory it is trained to. Further, we find that delta learning based on a semiempirical level of theory approximately halves the errors. We test the models on torsion energy profiles, relative conformational energies, geometric parameters, and relative tautomer errors.

High-Dimensional Neural Network Potential for Liquid Electrolyte Simulations

“…In comparison, the QRNN electrolyte potential very accurately predicts thermodynamic and transport properties across a subset of chemical compounds that are utilized in industrial liquid electrolytes. The dynamic charge model in the QRNN architecture can describe charge transfer and polarization effects, which are important when modeling ionic systems like electrolytes . Furthermore, it has been established that adding long-range dispersion and coulomb interactions are important for modeling neutral and ionic systems. , Overall, our results show promise for building a general liquid electrolyte model.…”

“…We trained a QRNN force field to a relatively small (∼360 K datapoints compared to ∼5 M for ANI-1×) data set of electrolyte clusters, including common carbonate solvents (EC, PC, VC, FEC, DMC, DEC, and EMC), Li + , and PF 6 − ions. To evaluate the model performance, we constructed an independently sampled test set of 2.5k cluster datapoints extracted from the production NPT trajectories of pure carbonate solvents and electrolyte mixtures.…”

“…As stated above, we used the previously reported QRNN architecture to train our electrolyte models using a locally modified copy of TorchANI open source software package. , Some small differences in the training protocol were used compared to the training of the direct learned ionic QRNN model previously reported . Specifically, throughout the active learning process, we trained on both energy and dipoles labels via multitask learning , while using early stopping and a 1% learning rate decay on plateau, with a patience of zero and a maximum of 500 epochs.…”

“…Unlike traditional force fields, these make few prior assumptions about the shape of the interatomic potentials. These models are more computationally efficient than quantum chemical methods and can achieve chemical accuracy relative to the method they are trained to reproduce. − Furthermore, MLFFs have shown to accurately reproduce both equilibrium and reactive regions of the PES, while traditional empirically fitted force fields have trouble with reactive systems without major modifications to the functional form and substantial parameterization. In essence, MLFFs are constructed by applying a regression algorithm to a highly flexible functional form such as Gaussian process regression, kernel rigid regression, or high-dimensional neural network potentials (HDNNPs) .…”

“…In this report, we construct an HDNNP, utilizing the recently reported charge recursive neural network architecture (QRNN), for liquid electrolyte simulations. The model was trained using an active learning approach. − In addition, the QRNN was trained exclusively to non-periodic DFT cluster data without any parametrization against experimental data (Scheme ), a strategy that has been proven to be successful for liquid water .…”