First principles reaction discovery: from the Schrodinger equation to experimental prediction for methane pyrolysis.

Our recent success in exploiting graphical processing units (GPUs) to accelerate quantum chemistry computations led to the development of the ab initio nanoreactor, a computational framework for automatic reaction discovery and kinetic model construction. In this work, we apply the ab initio nanoreactor to methane pyrolysis, from automatic reaction discovery to path refinement and kinetic modeling. Elementary reactions occurring during methane pyrolysis are revealed using GPU-accelerated ab initio molecular dynamics simulations. Subsequently, these reaction paths are refined at a higher level of theory with optimized reactant, product, and transition state geometries. Reaction rate coefficients are calculated by transition state theory based on the optimized reaction paths. The discovered reactions lead to a kinetic model with 53 species and 134 reactions, which is validated against experimental data and simulations using literature kinetic models. We highlight the advantage of leveraging local brute force and Monte Carlo sensitivity analysis approaches for efficient identification of important reactions. Both sensitivity approaches can further improve the accuracy of the methane pyrolysis kinetic model. The results in this work demonstrate the power of the ab initio nanoreactor framework for computationally affordable systematic reaction discovery and accurate kinetic modeling.

Single-Point Extrapolation to the Complete Basis Set Limit through Deep Learning.

Machine learning (ML) offers an attractive method for making predictions about molecular systems while circumventing the need to run expensive electronic structure calculations. Once trained on ab initio data, the promise of ML is to deliver accurate predictions of molecular properties that were previously computationally infeasible. In this work, we develop and train a graph neural network model to correct the basis set incompleteness error (BSIE) between a small and large basis set at the RHF and B3LYP levels of theory. Our results show that, when compared to fitting to the total potential, an ML model fitted to correct the BSIE is better at generalizing to systems not seen during training. We test this ability by training on single molecules while evaluating on molecular complexes. We also show that ensemble models yield better behaved potentials in situations where the training data is insufficient. However, even when only fitting to the BSIE, acceptable performance is only achieved when the training data sufficiently resemble the systems one wants to make predictions on. The test error of the final model trained to predict the difference between the cc-pVDZ and cc-pV5Z potential is 0.184 kcal/mol for the B3LYP density functional, and the ensemble model accurately reproduces the large basis set interaction energy curves on the S66x8 dataset.

ChemPix: automated recognition of hand-drawn hydrocarbon structures using deep learning.

Inputting molecules into chemistry software, such as quantum chemistry packages, currently requires domain expertise, expensive software and/or cumbersome procedures. Leveraging recent breakthroughs in machine learning, we develop ChemPix: an offline, hand-drawn hydrocarbon structure recognition tool designed to remove these barriers. A neural image captioning approach consisting of a convolutional neural network (CNN) encoder and a long short-term memory (LSTM) decoder learned a mapping from photographs of hand-drawn hydrocarbon structures to machine-readable SMILES representations. We generated a large auxiliary training dataset, based on RDKit molecular images, by combining image augmentation, image degradation and background addition. Additionally, a small dataset of ∼600 hand-drawn hydrocarbon chemical structures was crowd-sourced using a phone web application. These datasets were used to train the image-to-SMILES neural network with the goal of maximizing the hand-drawn hydrocarbon recognition accuracy. By forming a committee of the trained neural networks where each network casts one vote for the predicted molecule, we achieved a nearly 10 percentage point improvement of the molecule recognition accuracy and were able to assign a confidence value for the prediction based on the number of agreeing votes. The ensemble model achieved an accuracy of 76% on hand-drawn hydrocarbons, increasing to 86% if the top 3 predictions were considered.

The non-adiabatic nanoreactor: towards the automated discovery of photochemistry.

The ab initio nanoreactor has previously been introduced to automate reaction discovery for ground state chemistry. In this work, we present the nonadiabatic nanoreactor, an analogous framework for excited state reaction discovery. We automate the study of nonadiabatic decay mechanisms of molecules by probing the intersection seam between adiabatic electronic states with hyper-real metadynamics, sampling the branching plane for relevant conical intersections, and performing seam-constrained path searches. We illustrate the effectiveness of the nonadiabatic nanoreactor by applying it to benzene, a molecule with rich photochemistry and a wide array of photochemical products. Our study confirms the existence of several types of S0/S1 and S1/S2 conical intersections which mediate access to a variety of ground state stationary points. We elucidate the connections between conical intersection energy/topography and the resulting photoproduct distribution, which changes smoothly along seam space segments. The exploration is performed with minimal user input, and the protocol requires no previous knowledge of the photochemical behavior of a target molecule. We demonstrate that the nonadiabatic nanoreactor is a valuable tool for the automated exploration of photochemical reactions and their mechanisms.

TeraChem Cloud: A High-Performance Computing Service for Scalable Distributed GPU-Accelerated Electronic Structure Calculations.

The encapsulation and commoditization of electronic structure arise naturally as interoperability, and the use of nontraditional compute resources (e.g., new hardware accelerators, cloud computing) remains important for the computational chemistry community. We present TeraChem Cloud, a high-performance computing service (HPCS) that offers on-demand electronic structure calculations on both traditional HPC clusters and cloud-based hardware. The framework is designed using off-the-shelf web technologies and containerization to be extremely scalable and portable. Within the HPCS model, users can quickly develop new methods and algorithms in an interactive environment on their laptop while allowing TeraChem Cloud to distribute ab initio calculations across all available resources. This approach greatly increases the accessibility of hardware accelerators such as graphics processing units (GPUs) and flexibility for the development of new methods as additional electronic structure packages are integrated into the framework as alternative backends. Cost-performance analysis indicates that traditional nodes are the most cost-effective long-term solution, but commercial cloud providers offer cutting-edge hardware with competitive rates for short-term large-scale calculations. We demonstrate the power of the TeraChem Cloud framework by carrying out several showcase calculations, including the generation of 300,000 density functional theory energy and gradient evaluations on medium-sized organic molecules and reproducing 300 fs of nonadiabatic dynamics on the B800-B850 antenna complex in LH2, with the latter demonstration using over 50 Tesla V100 GPUs in a commercial cloud environment in 8 h for approximately $1250.

Geodesic interpolation for reaction pathways.

The development of high throughput reaction discovery methods such as the ab initio nanoreactor demands massive numbers of reaction rate calculations through the optimization of minimum energy reaction paths. These are often generated from interpolations between the reactant and product endpoint geometries. Unfortunately, straightforward interpolation in Cartesian coordinates often leads to poor approximations that lead to slow convergence. In this work, we reformulate the problem of interpolation between endpoint geometries as a search for the geodesic curve on a Riemannian manifold. We show that the perceived performance difference of interpolation methods in different coordinates is the result of an implicit metric change. Accounting for the metric explicitly allows us to obtain good results in Cartesian coordinates, bypassing the difficulties caused by redundant coordinates. Using only geometric information, we are able to generate paths from reactants to products which are remarkably close to the true minimum energy path. We show that these geodesic paths are excellent starting guesses for minimum energy path algorithms.

Large-Scale Functional Group Symmetry-Adapted Perturbation Theory on Graphical Processing Units.

Symmetry-adapted perturbation theory (SAPT) is a valuable method for analyzing intermolecular interactions. The functional group SAPT partition (F-SAPT) has been introduced to provide additional insight into the origins of noncovalent interactions. Until now, SAPT analysis has been too costly for large ligand-protein complexes where it could provide key insights for chemical modifications that might improve ligand binding. In this paper, we present a large-scale implementation of a variant of F-SAPT. Two pragmatic choices are made from the outset to render the problem tractable: (1) Ab initio computation of dispersion and exchange-dispersion is replaced with Grimme's empirical dispersion correction. (2) Basis sets with augmented functions are avoided to allow for efficient integral screening. These choices allow the F-SAPT analysis to be written largely in terms of Coulomb and exchange matrix builds which have been implemented efficiently on graphical processing units (GPUs). Our formulation of F-SAPT is routinely applicable to molecules with well over 3000 atoms and 25,000 basis functions and is particularly optimized for the case where one monomer is significantly larger than the other. This is demonstrated explicitly with results from F-SAPT analysis of the full indinavir @ HIV-II protease complex (PDB ID 1HSG ) in a polarized double-ζ basis.

A classical trajectory study of the photodissociation of T1 acetaldehyde: the transition from impulsive to statistical dynamics.

Previous experimental and theoretical studies of the radical dissociation channel of T(1) acetaldehyde show conflicting behavior in the HCO and CH(3) product distributions. To resolve these conflicts, a full-dimensional potential-energy surface for the dissociation of CH(3)CHO into HCO and CH(3) fragments over the barrier on the T(1) surface is developed based on RO-CCSD(T)/cc-pVTZ(DZ) ab initio calculations. 20,000 classical trajectories are calculated on this surface at each of five initial excess energies, spanning the excitation energies used in previous experimental studies, and translational, vibrational, and rotational distributions of the radical products are determined. For excess energies near the dissociation threshold, both the HCO and CH(3) products are vibrationally cold; there is a small amount of HCO rotational excitation and little CH(3) rotational excitation, and the reaction energy is partitioned dominantly (>90% at threshold) into relative translational motion. Close to threshold the HCO and CH(3) rotational distributions are symmetrically shaped, resembling a Gaussian function, in agreement with observed experimental HCO rotational distributions. As the excess energy increases the calculated HCO and CH(3) rotational distributions are observed to change from a Gaussian shape at threshold to one more resembling a Boltzmann distribution, a behavior also seen by various experimental groups. Thus the distribution of energy in these rotational degrees of freedom is observed to change from nonstatistical to apparently statistical, as excess energy increases. As the energy above threshold increases all the internal and external degrees of freedom are observed to gain population at a similar rate, broadly consistent with equipartitioning of the available energy at the transition state. These observations generally support the practice of separating the reaction dynamics into two reservoirs: an impulsive reservoir, fed by the exit channel dynamics, and a statistical reservoir, supported by the random distribution of excess energy above the barrier. The HCO rotation, however, is favored by approximately a factor of 3 over the statistical prediction. Thus, at sufficiently high excess energies, although the HCO rotational distribution may be considered statistical, the partitioning of energy into HCO rotation is not.

CH5+: chemistry's chameleon unmasked.

The nuclear vibrational wave function and zero-point vibrational energy of CH5(+) are calculated using quantum diffusion Monte Carlo techniques on an interpolated potential energy surface constructed from CCSD(T)/aug'-cc-pVTZ ab initio data. From this multidimensional wave function, the vibrationally averaged rotational constants and radial distribution functions for atom-atom distances within the molecule are constructed. It is found that the distributions of all 10 H-H distances are bimodal and identical. The radial distribution functions obtained for the five C-H distances are also identical, but unimodal. The three rotational constants were found to be 3.78, 3.80, and 3.83 cm(-1). These values indicate that the ground state of CH5(+) is significantly more symmetric than its global minimum energy structure. We conclude that the zero-point motion of CH5(+) renders all five protons equivalent in the ground state and precludes the assignment of a unique structure to the molecule.

On the extent of intramolecular hydrogen bonding in gas-phase and hydrated 1,2-ethanediol.

We investigate the quantum dynamical nature of hydrogen bonding in 1,2-ethanediol and monohydrated 1,2-ethanediol using different levels of ab initio theory. Global full-dimensional potential energy surfaces were constructed from PW91/cc-pVDZ, B3LYP/cc-pVDZ, and MP2/cc-pVDZ ab initio data for gas-phase and monohydrated 1,2-ethanediol, using a modified Shepard interpolation scheme. Zero-point energies and nuclear vibrational wave functions were calculated on these surfaces using the quantum diffusion Monte Carlo algorithm. The nature of intra- and intermolecular hydrogen bonding in these molecules was investigated by considering a ground-state nuclear vibrational wavefunction with reduced complete nuclear permutation and inversion (CNPI) symmetry. Separate wavefunction histograms were determined from the ground-state nuclear vibrational wavefunction by projection into bondlength coordinates. The O-H and O-O wavefunction histograms and vibrationally averaged distances were then used to probe the extent of intra- and intermolecular hydrogen bonding. From these data, we conclude that gas-phase ethanediol may possess a weak hydrogen bond, with a relatively short O-O distance but no detectable proton delocalization. Monohydrated ethanediol was found to exhibit no intramolecular hydrogen bonding but instead possessed two intermolecular hydrogen bonds, indicated by both shortening of the O-O distance and significant proton delocalization. The degree of proton delocalization and shortening of the vibrationally averaged O-O distance was found to be dependent on the ab initio method used to generate the potential energy surface (PES) data set.

Efficiency considerations in the construction of interpolated potential energy surfaces for the calculation of quantum observables by diffusion Monte Carlo.

A modified Shepard interpolation scheme is used to construct global potential energy surfaces (PES) in order to calculate quantum observables--vibrationally averaged internal coordinates, fully anharmonic zero-point energies and nuclear radial distribution functions--for a prototypical loosely bound molecular system, the water dimer. The efficiency of PES construction is examined with respect to (a) the method used to sample configurational space, (b) the method used to choose which points to add to the PES data set, and (c) the use of either a one- or two-part weight function. The most efficient method for constructing the PES is found to require a quantum sampling regime, a combination of both h-weight and rms methods for choosing data points and use of the two-part weight function in the interpolation. Using this regime, the quantum diffusion Monte Carlo zero-point energy converges to the exact result within addition of 50 data points. The vibrationally averaged O-O distance and O-O radial distribution function, however, converge more slowly and require addition of over 500 data points. The methods presented here are expected to be applicable to both other loosely bound complexes as well as tightly bound molecular species. When combined with high quality ab initio calculations, these methods should be able to accurately characterize the PES of such species.