--- title: "QVis: A Visual Analytics Tool for Exploring Noise and Errors in Quantum Computing Systems" authors: - Chad A. Steed - Junghoon Chae - Samudra Dasgupta - Travis S. Humble abstract: We present the preliminary design and results of QVis, a visual analytics tool for exploring quantum device performance data. QVis helps uncover temporal and multivariate variations in noise properties of quantum devices. We describe the implementations of these methods as well as applications to the analysis of a 127-qubit data set derived from the IBM washington processor over a 16-month period. Both human-interactive and semi-automated analytic methods are included to address requirements for visual exploration, thresholding, and clustering techniques. Our application of QVis to real-world scenarios demonstrates the ability to reveal noteworthy patterns in the behavior of the critical characterization metrics. summaryType: survey sourceStatus: null sources: - https://doi.org/10.1109/QCE57702.2023.10215 --- [Original paper (PDF)](https://doi.org/10.1109/QCE57702.2023.10215) ## Background and motivation QVis addresses the analysis of changing noise and error characteristics in quantum processors. The underlying problem is established: decoherence, imperfect gates, and measurement errors limit the reliability of quantum computation, and hardware characterization values can vary across qubits and over time. As processors grow, analysts must interpret many related time series rather than rely on a single device-level value or an isolated calibration snapshot. Figure 1 motivates this problem with the observed distribution of the $T_2$ coherence time for qubit 4 of IBM's Washington processor, showing substantial variation across the collection period. It is a data illustration of variability, not a QVis interface screenshot or evidence identifying its physical cause. The paper proposes an interactive environment for investigating these variations through established visualization and analysis techniques. Its methodological background includes focus-plus-context time-series exploration, Shneiderman's overview, zoom/filter, and details-on-demand strategy, parallel coordinates, and time-series clustering. The parallel-coordinate extensions draw on CrossVis, including statistical summaries and correlation indicators, while the web interface uses D3. The claimed need is for flexible exploration that combines broad temporal patterns with individual-qubit details and semi-automated guidance. The paper does not systematically compare existing quantum visualization tools or demonstrate that a particular earlier system fails on these tasks. Its contribution is an initial application and combination of these methods for quantum device characterization, rather than a new physical noise model or a new error-correction algorithm. ## Data and analytical scope The demonstration uses characterization data from the 127-qubit IBM Washington transmon processor collected between 1 January 2022 and 30 April 2023. The available data include state preparation and measurement error rates, gate error rates, gate durations, energy-relaxation time $T_1$, and coherence time $T_2$. The analyses reported in the paper concentrate on daily $T_1$ and $T_2$ measurements, for which larger values indicate more favorable information-storage behavior. These historical observations characterize the device during that collection period and do not establish its present performance. The interface lets users choose a metric and qubits of interest. Although the conclusion states that QVis is designed to support finer temporal scales, one- and two-qubit gate errors, and asymmetric readout error, the presented evidence is the daily relaxation and coherence data. The system supports exploratory interpretation of characterization measurements; the paper does not report circuit optimization, automatic correction of noise, or improved execution fidelity resulting from its use. ## Linked temporal views and interactions QVis combines three temporal panels, illustrated in Figure 2: a focus heatmap, a context heatmap, and detailed line charts. In the two heatmaps, horizontal position represents time and vertical position represents the selected metric value. Data are aggregated into two-dimensional time/value bins, and a sequential blue scale encodes the number of associated qubits, with darker blue indicating more qubits. This encoding reduces the overplotting that would result from drawing every qubit's entire time series at once, while preserving the distribution of values and its changes over time. For the displayed $T_2$ example, the authors identify a dense band around $50$ to $100\,\mu\mathrm{s}$ alongside substantial variation outside it. The context panel shows the broad temporal extent, with a gray rectangular brush indicating the time interval enlarged in the focus panel. Users can drag a new interval, adjust either endpoint, or pan the selection. The example data's daily granularity limits the minimum meaningful interval. Menu controls change the bin size, metric, and included qubits, and hovering over a bin exposes summaries such as the qubit count and median value. Thus, aggregation is adjustable and can be interrogated rather than being a fixed statistical reduction. Clicking a bin in the focus panel opens the corresponding qubits' raw time series in the detail panel. In Figure 2, a black outline marks the selected bin, and the detail panel displays Q26, Q30, Q36, and Q50. Hovering over the line chart reports their numerical values at a selected time. This interaction connects a pattern found in the aggregate display to the individual qubits responsible for it, although the overview itself necessarily suppresses individual trajectories. Figure 2 shows the implemented temporal interface: a brushed context interval determines the focus view, and selecting a focus bin reveals four qubits in the detailed panel. ## Comparing qubits with parallel coordinates The parallel-coordinate view in Figure 3 compares one selected metric across multiple qubits and dates. Each qubit has a vertical axis, an additional axis represents date, and each polyline represents one day's measurements across the qubit axes. This is therefore a comparison of qubit-specific values of the same metric in the illustrated view, rather than a display in which every axis is a different hardware metric. Box plots summarize each qubit's distribution, and correlation indicators support comparison with the selected qubit. Horizontal scrolling exposes qubits beyond those currently visible. Brushing a range on the Q114 axis highlights in blue the days when its $T_1$ falls in that range; unselected polylines remain faint gray. The date axis reveals that most selected high values occurred between January and mid-March 2022, with a few occurrences in June and August. The example also shows Q114 reaching a higher upper range than the other displayed qubits. For the selected days, the authors report that Q116 has the strongest illustrated correlation with Q114, $r = 0.45$. These observations identify dates and qubits for further investigation, but they do not establish why the values changed or demonstrate a causal relationship between qubits. ## Temporal clustering and findings QVis supplements manual exploration with $k$-means clustering of qubit time series for an individual metric. The method compares values at corresponding time points using Euclidean distance and minimizes within-cluster squared distances. The authors use equal-length records and choose this approach to avoid the additional computation of dynamic time warping. They set $k = 6$ by trial and error for the example, while allowing it to be changed according to analytical objectives. Consequently, the six groups are an exploratory configuration, not a validated optimal partition or an independently established taxonomy of device behavior. Figure 4 presents the $T_2$ clusters as small-multiple line charts. Each panel shows time horizontally and $T_2$ vertically, with semitransparent black trajectories for individual qubits and a red pointwise mean, called the barycenter in the paper. The panel title reports the cluster size. Cluster 1 initially peaks before settling near $30\,\mu\mathrm{s}$, whereas the larger clusters 2, 3, and 4 show more variable trajectories around different dominant levels, approximately $110$, $140$, and $75\,\mu\mathrm{s}$ respectively. Clusters 5 and 6 contain only a few qubits and are interpreted by the authors as outlying groups. The visualization makes shared temporal profiles and unusual groups available for investigation; small cluster size alone does not establish a defective device component or a particular noise mechanism. ## Contributions, evidence, and future work The main contribution is a preliminary visual analytics tool that joins aggregate temporal exploration, individual-qubit inspection, statistical comparison, and time-series clustering around quantum hardware characterization data. The paper provides concrete interface examples and findings from the Washington dataset, demonstrating how these methods can reveal temporal variation, correlations, and groups of qubits with similar behavior. The authors also report that quantum computing experts regard the initial system as promising, but provide no formal user-study protocol, participant results, controlled comparison, or quantitative assessment of analytical accuracy or efficiency. The evidence therefore supports the feasibility and illustrative usefulness of the workflow rather than a measured improvement over competing tools. The study is limited to one processor and primarily two characterization metrics, and it does not validate the clustering through alternative distance measures, cluster-quality scores, or downstream physical investigation. Pointwise Euclidean comparison treats measurements at the same time index as directly comparable, while the manually selected number of clusters affects the resulting grouping. The visualizations help formulate hypotheses, but identifying the causes of observed changes remains outside the demonstrated analysis. At publication, QVis was still under active development and was not yet publicly available online. The stated future work includes extending temporal and multivariate analysis, incorporating processor topology, and adding automated methods to find connections across performance metrics. The authors also envision a publicly accessible system that dynamically retrieves data from multiple quantum devices for monitoring and benchmarking. Those capabilities are presented as planned extensions rather than demonstrated features of the reported prototype.