Acceleration of Graph Neural Network-Based Prediction Models in Chemistry via Co-Design Optimization on Intelligence Processing Units

Atomic structure prediction and associated property calculations are the bedrock of chemical physics. Since high-fidelity ab initio modeling techniques for computing the structure and properties can be prohibitively expensive, this motivates the development of machine-learning (ML) models that make these predictions more efficiently. Training graph neural networks over large atomistic databases introduces unique computational challenges, such as the need to process millions of small graphs with variable size and support communication patterns that are distinct from learning over large graphs, such as social networks. We demonstrate a novel hardware-software codesign approach to scale up the training of atomistic graph neural networks (GNN) for structure and property prediction. First, to eliminate redundant computation and memory associated with alternative padding techniques and to improve throughput via minimizing communication, we formulate the effective coalescing of the batches of variable-size atomistic graphs as the bin packing problem and introduce a hardware-agnostic algorithm to pack these batches. In addition, we propose hardware-specific optimizations, including a planner and vectorization for the gather-scatter operations targeted for Graphcore's Intelligence Processing Unit (IPU), as well as model-specific optimizations such as merged communication collectives and optimized softplus. Putting these all together, we demonstrate the effectiveness of the proposed codesign approach by providing an implementation of a well-established atomistic GNN on the Graphcore IPUs. We evaluate the training performance on multiple atomistic graph databases with varying degrees of graph counts, sizes, and sparsity. We demonstrate that such a codesign approach can reduce the training time of atomistic GNNs and can improve their performance by up to 1.5× compared to the baseline implementation of the model on the IPUs. Additionally, we compare our IPU implementation with a Nvidia GPU-based implementation and show that our atomistic GNN implementation on the IPUs can run 1.8× faster on average compared to the execution time on the GPUs.

An open database of computed bulk ternary transition metal dichalcogenides

We present a dataset of structural relaxations of bulk ternary transition metal dichalcogenides (TMDs) computed via plane-wave density functional theory (DFT). We examined combinations of up to two chalcogenides with seven transition metals from groups 4-6 in octahedral (1T) or trigonal prismatic (2H) coordination. The full dataset consists of 672 unique stoichiometries, with a total of 50,337 individual configurations generated during structural relaxation. Our motivations for building this dataset are (1) to develop a training set for the generation of machine and deep learning models and (2) to obtain structural minima over a range of stoichiometries to support future electronic analyses. We provide the dataset as individual VASP xml files as well as all configurations encountered during relaxations collated into an ASE database with the corresponding total energy and atomic forces. In this report, we discuss the dataset in more detail and highlight interesting structural and electronic features of the relaxed structures.

Active sampling for neural network potentials: Accelerated simulations of shear-induced deformation in Cu-Ni multilayers

Neural network potentials (NNPs) can greatly accelerate atomistic simulations relative to ab initio methods, allowing one to sample a broader range of structural outcomes and transformation pathways. In this work, we demonstrate an active sampling algorithm that trains an NNP that is able to produce microstructural evolutions with accuracy comparable to those obtained by density functional theory, exemplified during structure optimizations for a model Cu-Ni multilayer system. We then use the NNP, in conjunction with a perturbation scheme, to stochastically sample structural and energetic changes caused by shear-induced deformation, demonstrating the range of possible intermixing and vacancy migration pathways that can be obtained as a result of the speedups provided by the NNP. The code to implement our active learning strategy and NNP-driven stochastic shear simulations is openly available at https://github.com/pnnl/Active-Sampling-for-Atomistic-Potentials.

Tracking the Chemical Evolution of Iodine Species Using Recurrent Neural Networks

We apply recurrent neural networks (RNNs) to predict the time evolution of the concentration profile of multiple species resulting from a set of interconnected chemical reactions. As a proof of concept of our approach, RNNs were trained on a synthetic dataset generated by solving the kinetic equations of a system of aqueous inorganic iodine reactions that can follow after nuclear reactor accidents. We examine the minimum dataset necessary to obtain accurate predictions and explore the ability of RNNs to interpolate and extrapolate when exposed to previously unseen data. We also investigate the limits of our RNN by evaluating the robustness of the training initialization on our dataset.

Degradable Polycaprolactone and Polylactide Homopolymer and Block Copolymer Brushes Prepared by Surface-Initiated Polymerization with Triazabicyclodecene and Zirconium Catalysts

Surface-initiated ring-opening polymerization (SI-ROP) of polycaprolactone (PCL) and polylactide (PLA) polymer brushes with controlled degradation rates were prepared on oxide substrates. PCL brushes were polymerized from hydroxyl-terminated monolayers utilizing triazabicyclodecene (TBD) as the polymerization catalyst. A consistent brush thickness of 40 nm could be achieved with a reproducible unique crystalline morphology. The organocatalyzed PCL brushes were chain extended using lactide in the presence of zirconium n-butoxide to successfully grow PCL/PLA block copolymer (PCL-b-PLA) brushes with a final thickness of 55 nm. The degradation properties of "grafted from" PCL brush and the PCL-b-PLA brush were compared to "grafted to" PCL brushes, and we observed that the brush density plays a major role in degradation kinetics. Solutions of methanol/water at pH 14 were used to better solvate the brushes and increase the kinetics of degradation. This framework enables a control of degradation that allows for the precise removal of these coatings.

Rapid electrochemical reduction of Ni(II) generates reactive monolayers for conjugated polymer brushes in one step

This article reports the development of a robust, one-step electrochemical technique to generate surface-bound conjugated polymers. The electrochemical reduction of arene diazonium salts at the surface of a gold electrode is used to generate tethered bromobenzene monolayers quickly. The oxidative addition of reactive Ni(0) across the aryl halide bond is achieved in situ through a concerted electrochemical reduction of Ni(dppp)Cl2. This technique limits the diffusion of Ni(0) species away from the surface and overcomes the need for solution deposition techniques which often require multiple steps that result in a loss of surface coverage. With this electrochemical technique, the formation of the reactive monolayer resulted in a surface coverage of 1.29 × 10(14) molecules/cm(2), which is a 6-fold increase over previously reported results using solution deposition techniques.

On the Role of Disproportionation Energy in Kumada Catalyst-Transfer Polycondensation

Kumada catalyst-transfer polycondensation (KCTP) is an effective method for the controlled polymerization of conjugated polymers. Nevertheless, side reactions leading to early termination and unwanted chain coupling cause deviations from the target molecular weight, along with increasing polydispersity and end group variation. The departure from the KCTP cycle stems from a disproportionation reaction that leads to experimentally observed side products. The disproportionation energies for a series of nickel-based initiators containing bidentate phosphino attendant ligands were computed using density functional theory at the B3LYP/DZP level. The initiator was found to be less favorable toward disproportionation by 0.5 kcal mol^-1 when ligated by 1,3-bis(diphenylphosphino)propane (dppp) rather than 1,2-bis(diphenylphosphino)ethane (dppe). Trends in disproportionation energy (E_disp) with a variety of bidentate phosphine ligands match experimental observations of decreased polymerization control. Theoretical E_disp values can thus be used to predict the likelihood of disproportionation in cross-coupling reactions and, therefore, aid in catalyst design.