Feature tracking in time-varying scalar fields is a fundamental task in scientific computing. Topological descriptors, which summarize important features of data, have proved to be viable tools to facilitate this task. The merge tree is a topological descriptor that captures the connectivity behaviors of the sub- or superlevel sets of a scalar field. Edit distances between merge trees play a vital role in effective temporal data tracking. Existing methods to compute them fall into two main classes, namely whether they are dependent or independent of the branch decomposition. These two classes represent the most prominent approaches for producing tracking results. In this paper, we compare four different merge tree edit distance-based methods for feature tracking. We demonstrate that these methods yield distinct results with both analytical and real-world data sets. Furthermore, we investigate how these results vary and identify the factors that influence them. Our experiments reveal significant differences in tracked features over time, even among those produced by techniques within the same category.

Human preferences are widely used to align large language models (LLMs) through methods such as reinforcement learning from human feedback (RLHF). However, the current user interfaces require annotators to compare text paragraphs, which is cognitively challenging when the texts are long or unfamiliar. This paper contributes by studying the decomposition principle as an approach to improving the quality of human feedback for LLM alignment. This approach breaks down the text into individual claims instead of directly comparing two long-form text responses. Based on the principle, we build a novel user interface DxHF. It enhances the comparison process by showing decomposed claims, visually encoding the relevance of claims to the conversation and linking similar claims. This allows users to skim through key information and identify differences for better and quicker judgment. Our technical evaluation shows evidence that decomposition generally improves feedback accuracy regarding the ground truth, particularly for users with uncertainty. A crowdsourcing study with 160 participants indicates that using DxHF improves feedback accuracy by an average of 5%, although it increases the average feedback time by 18 seconds. Notably, accuracy is significantly higher in situations where users have less certainty. The finding of the study highlights the potential of HCI as an effective method for improving human-AI alignment.

Reinforcement learning from human feedback (RLHF) has emerged as a key enabling technology for aligning AI behaviour with human preferences. The traditional way to collect data in RLHF is via pairwise comparisons: human raters are asked to indicate which one of two samples they prefer. We present an interactive visualisation that better exploits the human visual ability to compare and explore whole groups of samples. The interface is comprised of two linked views: 1) an exploration view showing a contextual overview of all sampled behaviours organised in a hierarchical clustering structure; and 2) a comparison view displaying two selected groups of behaviours for user queries. Users can efficiently explore large sets of behaviours by iterating between these two views. Additionally, we devised an active learning approach suggesting groups for comparison. As shown by our evaluation in six simulated robotics tasks, our approach increases the final rewards by 69.34%. It leads to lower error rates and better policies. We open-source the code that can be easily integrated into the RLHF training loop, supporting research on human–AI alignment.

Wall-bounded turbulent flows as those occurring in transportation (e.g. aviation) or industrial applications (e.g turbomachinery), are usually subjected to pressure gradients (PGs). The presence of such PGs affects greatly the development and physics of the turbulent boundary layer (TBL), making it an open research area. An important phenomena associated with the presence of strong adverse PGs (APGs) as appearing in wings, is the separation of the boundary layer, which can lead to stall.

The transition from laminar to turbulent flow is a long-standing research subject in the field of fluid mechanics. A basic necessity for such studies is a distinction between laminar and turbulent flow. In particular, a binary segmentation is desired where laminar and turbulent regions behave consistently over time. Previous works in this regard yield inconsistent results, or are restricted to 2D manifolds thereby neglecting the three-dimensional nature of the problem. In this paper, we present a novel use of feature-based methods to segmenting a 3D time-dependent flow into regions of laminar and turbulent behavior. It is based on the extraction of local extrema from a scalar field such as spanwise velocity. It turns out that the existence of many extrema in a region is a good indicator for turbulence. We derive a density function from the extracted extrema using a Kernel Density Estimate (KDE) and threshold it to achieve a binary segmentation into laminar and turbulent regions. We use an in-situ processing approach for data analysis during the simulation run. The two core components of our method exhibit drastically different performance characteristics: the extraction of extrema is embarrassingly parallel, while the KDE is more time-consuming. Hence, we decouple our algorithmic components to achieve a better overall system performance. Our method shows consistent results and enables the domain scientists to study the three-dimensional aspects of the laminar-turbulent transition that were difficult to assess before.

In reinforcement learning, tuning reward weights in the reward function is necessary to align behavior with user preferences. However, current approaches, which use pairwise comparisons for preference elicitation, are inefficient, because they miss much of the human ability to explore and judge groups of candidate solutions. The paper presents a novel visualization-based approach that better exploits the user's ability to quickly recognize interesting directions for reward tuning. It breaks down the tuning problem by using the visual information-seeking principle: overview first, zoom and filter, then details-on-demand. Following this principle, we built a visualization system comprising two interactively linked views: 1) an embedding view showing a contextual overview of all sampled behaviors and 2) a sample view displaying selected behaviors and visualizations of the detailed time-series data. A user can efficiently explore large sets of samples by iterating between these two views. The paper demonstrates that the proposed approach is capable of tuning rewards for challenging behaviors. The simulation-based evaluation shows that the system can reach optimal solutions with fewer queries relative to baselines.

The Morse-Smale complex is a standard tool in visual data analysis. The classic definition is based on a continuous view of the gradient of a scalar function where its zeros are the critical points. These points are connected via gradient curves and surfaces emanating from saddle points, known as separatrices. In a discrete setting, the Morse-Smale complex is commonly extracted by constructing a combinatorial gradient assuming the steepest descent direction. Previous works have shown that this method results in a geometric embedding of the separatrices that can be fundamentally different from those in the continuous case. To achieve a similar embedding, different approaches for constructing a combinatorial gradient were proposed. In this paper, we show that these approaches generate a different topology, i.e., the connectivity between critical points changes. Additionally, we demonstrate that the steepest descent method can compute topologically and geometrically accurate Morse-Smale complexes when applied to certain types of grids. Based on these observations, we suggest a method to attain both geometric and topological accuracy for the Morse-Smale complex of data sampled on a uniform grid.

Scatterplots are commonly used in various contexts, from scientific publications to infographics for the general public. However, not everyone is able to read them, and even experts may struggle to notice some important information such as overlapping clusters or temporal changes. To address these issues, a computational approach for annotating scatterplots has been developed. This approach involves various forms of annotation, including drawing lines to show correlations, circling areas to show clusters, and indicating movement with arrows. The approach is based on a study that identified common annotation strategies used by people to annotate scatterplots. These strategies are distilled into an automated method for generating graphical annotations on scatterplots. The method involves a problem formulation using a Markov Decision Process and a model for making annotation decisions. The model generates step-by-step graphical annotations by analyzing data insights and observing the chart. The final result conveys a narrative that is easy to understand and allows for the conveyance of temporal changes in the data. The study results suggest that the method can generate understandable and functional annotations that are comparable to those created by human experts. This approach can potentially reduce the time and effort required to read scatterplots, making it a useful tool for data visualization novices.

We detail our developments in the high-fidelity spectral-element code Neko that are essential for unprecedented large-scale direct numerical simulations of fully developed turbulence. Major inno- vations are modular multi-backend design enabling performance portability across a wide range of GPUs and CPUs, a GPU-optimized preconditioner with task overlapping for the pressure-Poisson equation and in-situ data compression. We carry out initial runs of Rayleigh–Bénard Convection (RBC) at extreme scale on the LUMI and Leonardo supercomputers. We show how Neko is able to strongly scale to 16,384 GPUs and obtain results that are not pos- sible without careful consideration and optimization of the entire simulation workflow. These developments in Neko will help resolv- ing the long-standing question regarding the ultimate regime in RBC. 

Dendritic spines are submicron, subcellular compartments whose shape is defined by actin filaments and associated proteins. Accurately mapping the cytoskeleton is a challenge, given the small size of its components. It remains unclear whether the actin-associated structures analyzed in dendritic spines of neurons in vitro apply to dendritic spines of intact, mature neurons in situ. Here, we combined advanced preparative methods with multitilt serial section electron microscopy (EM) tomography and computational analysis to reveal the full three-dimensional (3D) internal architecture of spines in the intact brains of male mice at nanometer resolution. We compared hippocampal (CA1) pyramidal cells and cerebellar Purkinje cells in terms of the length distribution and connectivity of filaments, their branching-angles and absolute orientations, and the elementary loops formed by the network. Despite differences in shape and size across spines and between spine heads and necks, the internal organization was remarkably similar in both neuron types and largely homogeneous throughout the spine volume. In the tortuous mesh of highly branched and interconnected filaments, branches exhibited no preferred orientation except in the immediate vicinity of the cell membrane. We found that new filaments preferentially split off from the convex side of a bending filament, consistent with the behavior of Arp2/3-mediated branching of actin under mechanical deformation. Based on the quantitative analysis, the spine cytoskeleton is likely subject to considerable mechanical force in situ.