My Model is Unfair, Do People Even Care? Visual Design Affects Trust and Perceived Bias in Machine Learning.

Machine learning technology has become ubiquitous, but, unfortunately, often exhibits bias. As a consequence, disparate stakeholders need to interact with and make informed decisions about using machine learning models in everyday systems. Visualization technology can support stakeholders in understanding and evaluating trade-offs between, for example, accuracy and fairness of models. This paper aims to empirically answer "Can visualization design choices affect a stakeholder's perception of model bias, trust in a model, and willingness to adopt a model?" Through a series of controlled, crowd-sourced experiments with more than 1,500 participants, we identify a set of strategies people follow in deciding which models to trust. Our results show that men and women prioritize fairness and performance differently and that visual design choices significantly affect that prioritization. For example, women trust fairer models more often than men do, participants value fairness more when it is explained using text than as a bar chart, and being explicitly told a model is biased has a bigger impact than showing past biased performance. We test the generalizability of our results by comparing the effect of multiple textual and visual design choices and offer potential explanations of the cognitive mechanisms behind the difference in fairness perception and trust. Our research guides design considerations to support future work developing visualization systems for machine learning.

Professional Differences: A Comparative Study of Visualization Task Performance and Spatial Ability Across Disciplines.

Problem-driven visualization work is rooted in deeply understanding the data, actors, processes, and workflows of a target domain. However, an individual's personality traits and cognitive abilities may also influence visualization use. Diverse user needs and abilities raise natural questions for specificity in visualization design: Could individuals from different domains exhibit performance differences when using visualizations? Are any systematic variations related to their cognitive abilities? This study bridges domain-specific perspectives on visualization design with those provided by cognition and perception. We measure variations in visualization task performance across chemistry, computer science, and education, and relate these differences to variations in spatial ability. We conducted an online study with over 60 domain experts consisting of tasks related to pie charts, isocontour plots, and 3D scatterplots, and grounded by a well-documented spatial ability test. Task performance (correctness) varied with profession across more complex visualizations (isocontour plots and scatterplots), but not pie charts, a comparatively common visualization. We found that correctness correlates with spatial ability, and the professions differ in terms of spatial ability. These results indicate that domains differ not only in the specifics of their data and tasks, but also in terms of how effectively their constituent members engage with visualizations and their cognitive traits. Analyzing participants' confidence and strategy comments suggests that focusing on performance neglects important nuances, such as differing approaches to engage with even common visualizations and potential skill transference. Our findings offer a fresh perspective on discipline-specific visualization with specific recommendations to help guide visualization design that celebrates the uniqueness of the disciplines and individuals we seek to serve.

Quasi-continuous melting of model polymer monolayers prompts reinterpretation of polymer melting.

Condensed matter textbooks teach us that melting cannot be continuous and indeed experience, including with polymers and other long-chain compounds, tells us that it is a strongly first-order transition. However, here we report nearly continuous melting of monolayers of ultralong n-alkane C390H782 on graphite, observed by AFM and reproduced by mean-field theory and MD simulation. On heating, the crystal-melt interface moves steadily and reversibly from chain ends inward. Remarkably, the final melting point is 80 K above that of the bulk, and equilibrium crystallinity decreases continuously from ~100% to <50% prior to final melting. We show that the similarity in melting behavior of polymers and non-polymers is coincidental. In the bulk, the intermediate melting stages of long-chain crystals are forbidden by steric overcrowding at the crystal-liquid interface. However, there is no crowding in a monolayer as chain segments can escape to the third dimension.

Property Decoupling across the Embryonic Nucleus-Melt Interface during Polymer Crystal Nucleation.

Spatial distributions are presented that quantitatively capture how polymer properties (e.g., segment alignment, density, and potential energy) vary with distance from nascent polymer crystals (nuclei) in prototypical polyethylene melts. It is revealed that the spatial extent of nuclei and their interfaces is metric-dependent as is the extent to which nucleus interiors are solid-like. As distance from a nucleus increases, some properties, such as density, decay to melt-like behavior more rapidly than polymer segment alignment, indicating that a polymer nucleus resides in a nematic-like droplet. This nematic-like droplet region coincides with enhanced formation of ordered polymer segments that are not part of the nucleus. It is more favorable to find nonconstituent ordered polymer segments near a nucleus than in the surrounding metastable melt, pointing to the possibility of one nucleus inducing the formation of other nuclei. In this vein, there is also a second region of enhanced ordering that lies along the nematic director of a nucleus, but beyond its nematic droplet and fold regions. These results indicate that crystal stacking, a key characteristic of lamellae in semicrystalline polymeric materials, begins to emerge during the earliest stages of polymer crystallization (i.e., crystal nucleation). More generally, the findings of this study provide a conceptual bridge between polymer crystal nucleation under nonflow and flow conditions and are used to rationalize previous results.

Coexistence of Lipid Phases Stabilizes Interstitial Water in the Outer Layer of Mammalian Skin.

The lipid matrix in the outer layer of mammalian skin, the stratum corneum, has been previously investigated by multiple biophysical techniques aimed at identifying hydrophilic and lipophilic pathways of permeation. Although consensus is developing over the microscopic structure of the lipid matrix, no molecular-resolution model describes the permeability of all chemical species simultaneously. Using molecular dynamics simulations of a model mixture of skin lipids, the self-assembly of the lipid matrix lamellae has been studied. At higher humidity, the resulting lamellar phase is maintained by partitioning excess water into isolated droplets of controlled size and spatial distribution. The droplets may fuse together to form intralamellar water channels, thereby providing a pathway for the permeation of hydrophilic species. These results reconcile competing data on the outer skin's structure and broaden the scope of molecular-based methods to improve the safety of topical products and to advance transdermal drug delivery.

Monodisperse Polymer Melts Crystallize via Structurally Polydisperse Nanoscale Clusters: Insights from Polyethylene.

This study demonstrates that monodisperse entangled polymer melts crystallize via the formation of nanoscale nascent polymer crystals (i.e., nuclei) that exhibit substantial variability in terms of their constituent crystalline polymer chain segments (stems). More specifically, large-scale coarse-grain molecular simulations are used to quantify the evolution of stem length distributions and their properties during the formation of polymer nuclei in supercooled prototypical polyethylene melts. Stems can adopt a range of lengths within an individual nucleus (e.g., ∼1-10 nm) while two nuclei of comparable size can have markedly different stem distributions. As such, the attainment of chemically monodisperse polymer specimens is not sufficient to achieve physical uniformity and consistency. Furthermore, stem length distributions and their evolution indicate that polymer crystal nucleation (i.e., the initial emergence of a nascent crystal) is phenomenologically distinct from crystal growth. These results highlight that the tailoring of polymeric materials requires strategies for controlling polymer crystal nucleation and growth at the nanoscale.

Design by Immersion: A Transdisciplinary Approach to Problem-Driven Visualizations.

While previous work exists on how to conduct and disseminate insights from problem-driven visualization projects and design studies, the literature does not address how to accomplish these goals in transdisciplinary teams in ways that advance all disciplines involved. In this paper we introduce and define a new methodological paradigm we call design by immersion, which provides an alternative perspective on problem-driven visualization work. Design by immersion embeds transdisciplinary experiences at the center of the visualization process by having visualization researchers participate in the work of the target domain (or domain experts participate in visualization research). Based on our own combined experiences of working on cross-disciplinary, problem-driven visualization projects, we present six case studies that expose the opportunities that design by immersion enables, including (1) exploring new domain-inspired visualization design spaces, (2) enriching domain understanding through personal experiences, and (3) building strong transdisciplinary relationships. Furthermore, we illustrate how the process of design by immersion opens up a diverse set of design activities that can be combined in different ways depending on the type of collaboration, project, and goals. Finally, we discuss the challenges and potential pitfalls of design by immersion.

Divining the shape of nascent polymer crystal nuclei.

We demonstrate that nascent polymer crystals (i.e., nuclei) are anisotropic entities with neither spherical nor cylindrical geometry, in contrast to previous assumptions. In fact, cylindrical, spherical, and other high symmetry geometries are thermodynamically unfavorable. Moreover, postcritical transitions are necessary to achieve the lamellae that ultimately arise during the crystallization of semicrystalline polymers. We also highlight how inaccurate treatments of polymer nucleation can lead to substantial errors (e.g., orders of magnitude discrepancies in predicted nucleation rates). These insights are based on quantitative analysis of over four million crystal clusters from the crystallization of prototypical entangled polyethylene melts. New comprehensive bottom-up models are needed to capture polymer nucleation.

A coarse-grain model for entangled polyethylene melts and polyethylene crystallization.

The Shinoda-DeVane-Klein (SDK) model is herein demonstrated to be a viable coarse-grain model for performing molecular simulations of polyethylene (PE), affording new opportunities to advance molecular-level, scientific understanding of PE materials and processes. Both structural and dynamical properties of entangled PE melts are captured by the SDK model, which also recovers important aspects of PE crystallization phenomenology. Importantly, the SDK model can be used to represent a variety of materials beyond PE and has a simple functional form, making it unique among coarse-grain PE models. This study expands the suite of tools for studying PE in silico and paves the way for future work probing PE and PE-based composites at the molecular level.

Characterizing key features in the formation of ice and gas hydrate systems.

Crystallization in liquids is critical to a range of important processes occurring in physics, chemistry and life sciences. In this article, we review our efforts towards understanding the crystallization mechanisms, where we focus on theoretical modelling and molecular simulations applied to ice and gas hydrate systems. We discuss the order parameters used to characterize molecular ordering processes and how different order parameters offer different perspectives of the underlying mechanisms of crystallization. With extensive simulations of water and gas hydrate systems, we have revealed unexpected defective structures and demonstrated their important roles in crystallization processes. Nucleation of gas hydrates can in most cases be characterized to take place in a two-step mechanism where the nucleation occurs via intermediate metastable precursors, which gradually reorganizes to a stable crystalline phase. We have examined the potential energy landscapes explored by systems during nucleation, and have shown that these landscapes are rugged and funnel-shaped. These insights provide a new framework for understanding nucleation phenomena that has not been addressed in classical nucleation theory. This article is part of the theme issue 'The physics and chemistry of ice: scaffolding across scales, from the viability of life to the formation of planets'.

Polymer nucleation under high-driving force, long-chain conditions: Heat release and the separation of time scales.

This study reveals important features of polymer crystal formation at high-driving forces in entangled polymer melts based on simulations of polyethylene. First and in contrast to small-molecule crystallization, the heat released during polymer crystallization does not appreciably influence structural details of early-stage, crystalline clusters (crystal nuclei). Second, early-stage polymer crystallization (crystal nucleation) can occur without substantial chain-level relaxation and conformational changes. This study's results indicate that local structures and environments guide crystal nucleation in entangled polymer melts under high-driving force conditions. Given that such conditions are often used to process polyethylene, local structures and the separation of time scales associated with crystallization and chain-level processes are anticipated to be of substantial importance to processing strategies. This study highlights new research directions for understanding polymer crystallization.

Does Local Structure Bias How a Crystal Nucleus Evolves?

The broad scientific and technological importance of crystallization has led to significant research probing and rationalizing crystal nucleation processes. Previous work has generally neglected the possibility of the molecular-level dynamics of individual crystal nuclei coupling to local structures. However, recent experimental work has conjectured that this can occur. Therefore, to address a deficiency in scientific understanding of crystallization, we have probed the nucleation of prototypical single and multicomponent crystals (specifically, ice and mixed gas hydrates). We establish that local structures can bias the evolution of nascent crystal phases on a nanosecond time scale by, for example, promoting the appearance or disappearance of specific crystal motifs and thus reveal a new facet of crystallization behavior. Moreover, we demonstrate structural biases are likely present during crystallization processes beyond ice and gas hydrate formation. Structurally biased dynamics are a lens for understanding existing computational and experimental results while pointing to future opportunities.

Unraveling Mixed Hydrate Formation: Microscopic Insights into Early Stage Behavior.

The molecular-level details of mixed hydrate nucleation remain unclear despite the broad implications of this process for a variety of scientific domains. Through analysis of mixed hydrate nucleation in a prototypical CH4/H2S/H2O system, we demonstrate that high-level kinetic similarities between mixed hydrate systems and corresponding pure hydrate systems are not a reliable basis for estimating the composition of early stage mixed hydrate nuclei. Moreover, we show that solution compositions prior to and during nucleation are not necessarily effective proxies for the composition of early stage mixed hydrate nuclei. Rather, microscopic details, (e.g., guest-host interactions and previously neglected cage types) apparently play key roles in determining early stage behavior of mixed hydrates. This work thus provides key foundational concepts and insights for understanding mixed hydrate nucleation.

Evidence from mixed hydrate nucleation for a funnel model of crystallization.

The molecular-level details of crystallization remain unclear for many systems. Previous work has speculated on the phenomenological similarities between molecular crystallization and protein folding. Here we demonstrate that molecular crystallization can involve funnel-shaped potential energy landscapes through a detailed analysis of mixed gas hydrate nucleation, a prototypical multicomponent crystallization process. Through this, we contribute both: (i) a powerful conceptual framework for exploring and rationalizing molecular crystallization, and (ii) an explanation of phenomenological similarities between protein folding and crystallization. Such funnel-shaped potential energy landscapes may be typical of broad classes of molecular ordering processes, and can provide a new perspective for both studying and understanding these processes.

Mechanism of O((3)P) formation from a hydroxyl radical pair in aqueous solution.

The reaction mechanism for the rapid formation of a triplet oxygen atom, O((3)P), from a pair of triplet-state hydroxyl radicals in liquid water is explored utilizing extensive Car-Parrinello MD simulations and advanced visualization techniques. The local solvation structures, the evolution of atomic charges, atomic separations, spin densities, electron localization functions, and frontier molecular orbitals, as well as free energy profiles, evidence that the reaction proceeds through a hybrid (hydrogen atom transfer and electron-proton transfer) and hemibond-assisted reaction mechanism. A benchmarking study utilizing high-level ab initio calculations to examine the interactions of a hydroxyl radical pair in the gas phase and the influence of a hemibonded water is also provided. The results presented here should serve as a foundation for further experimental and theoretical studies aimed at better understanding the role and potential applications of the triplet oxygen atom as a potent reactive oxygen species.

Algorithms for GPU-based molecular dynamics simulations of complex fluids: Applications to water, mixtures, and liquid crystals.

A custom code for molecular dynamics simulations has been designed to run on CUDA-enabled NVIDIA graphics processing units (GPUs). The double-precision code simulates multicomponent fluids, with intramolecular and intermolecular forces, coarse-grained and atomistic models, holonomic constraints, Nosé-Hoover thermostats, and the generation of distribution functions. Algorithms to compute Lennard-Jones and Gay-Berne interactions, and the electrostatic force using Ewald summations, are discussed. A neighbor list is introduced to improve scaling with respect to system size. Three test systems are examined: SPC/E water; an n-hexane/2-propanol mixture; and a liquid crystal mesogen, 2-(4-butyloxyphenyl)-5-octyloxypyrimidine. Code performance is analyzed for each system. With one GPU, a 33-119 fold increase in performance is achieved compared with the serial code while the use of two GPUs leads to a 69-287 fold improvement and three GPUs yield a 101-377 fold speedup.

Aqueous production of oxygen atoms from hydroxyl radicals.

The behavior of the hydroxyl radical (OH*) in solution is significant to a broad range of scientific and technological fields. OH* is considered a highly reactive, short-lived species and previous studies have neglected the possibility of encounters of two OH* in solution. However, these encounters may be nonnegligible in environments with elevated local OH* concentrations, such as under many in vivo processes and within nuclear infrastructure. High concentrations of OH* in vivo are considered to be very dangerous; OH* has been related to many ailments ranging from cancer to Alzheimer's disease. Here we probe details of the reactions and interactions that can occur between two OH* in water by utilizing Car-Parrinello molecular dynamics simulations and advanced visualization techniques. The recombination reaction to form hydrogen peroxide is confirmed for the singlet electronic state. In contrast, the triplet state yields an oxygen atom, O(aq). This species has been previously detected in experimental water-radiolysis studies, but its origin could not be determined. O(aq) is a much more potent biradical than its parent OH* and its presence can impact many in vivo processes. This study also reveals that the hemibonded interaction plays key role in the behavior of OH*(aq). Our findings have major implications to the scientific understanding of the impacts of high local OH* concentrations, such during oxidative stress and in aging processes. Given its importance, this study will form the basis of further experimental and theoretical investigations exploring the role of O(aq) in a number of contexts.