Search for: All records

We develop a measure for evaluating the performance of generative networks given two sets of images. A popular performance measure currently used to do this is the Fréchet Inception Distance (FID). FID assumes that images featurized using the penultimate layer of Inception-v3 follow a Gaussian distribution, an assumption which cannot be violated if we wish to use FID as a metric. However, we show that Inception-v3 features of the ImageNet dataset are not Gaussian; in particular, every single marginal is not Gaussian. To remedy this problem, we model the featurized images using Gaussian mixture models (GMMs) and compute the 2-Wasserstein distance restricted to GMMs. We define a performance measure, which we call WaM, on two sets of images by using Inception-v3 (or another classifier) to featurize the images, estimate two GMMs, and use the restricted 2-Wasserstein distance to compare the GMMs. We experimentally show the advantages of WaM over FID, including how FID is more sensitive than WaM to imperceptible image perturbations. By modelling the non-Gaussian features obtained from Inception-v3 as GMMs and using a GMM metric, we can more accurately evaluate generative network performance. 

The objective of this work is to predict the morphology and material properties of crosslinking polymers used in aerospace applications. We extend the open-source dybond plugin for HOOMD-Blue to implement a new coarse-grained model of reacting epoxy thermosets and use the 44DDS/DGEBA/PES system as a case study for calibration and validation. We parameterize the coarse-grained model from atomistic solubility data, calibrate reaction dynamics against experiments, and check for size-dependent artifacts. We validate model predictions by comparing glass transition temperatures measurements at arbitrary degree of cure, gel-points, and morphology predictions against experiments. We demonstrate for the first time in molecular simulations the cure-path dependence of toughened thermoset morphologies. 

We develop an optimized force-field for poly(3-hexylthiophene) (P3HT) and demonstrate its utility for predicting thermodynamic self-assembly. In particular, we consider short oligomer chains, model electrostatics and solvent implicitly, and coarsely model solvent evaporation. We quantify the performance of our model to determine what the optimal system sizes are for exploring self-assembly at combinations of state variables. We perform molecular dynamics simulations to predict the self-assembly of P3HT at ∼350 combinations of temperature and solvent quality. Our structural calculations predict that the highest degrees of order are obtained with good solvents just below the melting temperature. We find our model produces the most accurate structural predictions to date, as measured by agreement with grazing incident X-ray scattering experiments. 

The predictive capabilities of computational materials science today derive from overlapping advances in simulation tools, modeling techniques, and best practices. We outline this ecosystem of molecular simulations by explaining how important contributions in each of these areas have fed into each other. The combined output of these tools, techniques, and practices is the ability for researchers to advance understanding by efficiently combining simple models with powerful software. As specific examples, we show how the prediction of organic photovoltaic morphologies have improved by orders of magnitude over the last decade, and how the processing of reacting epoxy thermosets can now be investigated with million-particle models. We discuss these two materials systems and the training of materials simulators through the lens of cognitive load theory. For students, the broad view of ecosystem components should facilitate understanding how the key parts relate to each other first, followed by targeted exploration. In this way, the paper is organized in loose analogy to a coarse-grained model: The main components provide basic framing and accelerated sampling from which deeper research is better contextualized. For mentors, this paper is organized to provide a snapshot in time of the current simulation ecosystem and an on-ramp for simulation experts into the literature on pedagogical practice. 

We report the results of the COVID Moonshot, a fully open-science, crowdsourced, and structure-enabled drug discovery campaign targeting the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) main protease. We discovered a noncovalent, nonpeptidic inhibitor scaffold with lead-like properties that is differentiated from current main protease inhibitors. Our approach leveraged crowdsourcing, machine learning, exascale molecular simulations, and high-throughput structural biology and chemistry. We generated a detailed map of the structural plasticity of the SARS-CoV-2 main protease, extensive structure-activity relationships for multiple chemotypes, and a wealth of biochemical activity data. All compound designs (>18,000 designs), crystallographic data (>490 ligand-bound x-ray structures), assay data (>10,000 measurements), and synthesized molecules (>2400 compounds) for this campaign were shared rapidly and openly, creating a rich, open, and intellectual property–free knowledge base for future anticoronavirus drug discovery.