--- title: Visual Analytics of Performance of Quantum Computing Systems and Circuit Optimization authors: - Junghoon Chae - Chad A. Steed - Travis S. Humble abstract: Driven by potential exponential speedups in business, security, and scientific scenarios, interest in quantum computing is surging. This interest feeds the development of quantum computing hardware, but several challenges arise in optimizing application performance for hardware metrics (e.g., qubit coherence and gate fidelity). In this work, we describe a visual analytics approach for analyzing the performance properties of quantum devices and quantum circuit optimization. Our approach allows users to explore spatial and temporal patterns in quantum device performance data and it computes similarities and variances in key performance metrics. Detailed analysis of the error properties characterizing individual qubits is also supported. We also describe a method for visualizing the optimization of quantum circuits. The resulting visualization tool allows researchers to design more efficient quantum algorithms and applications by increasing the interpretability of quantum computations. summaryType: survey sourceStatus: null sources: - https://doi.org/10.1109/ISVLSI61997.2024.00116 --- # Visual Analytics of Performance of Quantum Computing Systems and Circuit Optimization [Read the original paper](https://doi.org/10.1109/ISVLSI61997.2024.00116) ## Background and motivation Quantum circuit developers need to understand both the reliability of a device and the changes introduced when a circuit is compiled for its hardware constraints. Calibration properties vary between qubits and over time, so a single device-wide average or one calibration snapshot can conceal important differences. The paper illustrates this variability with a distribution of coherence times for qubit 4 of IBM's Washington processor in Figure 1. Circuit optimization creates a related interpretation problem: transformations such as gate reordering, fusion, elimination, and qubit mapping change circuit structure, and developers need to examine whether those changes reduce the resources required by their computation. The authors present QVis as a visual analytics environment for exploring these two problems through hardware-characterization views and a circuit-optimization interface. Noise analysis and circuit optimization are existing problems, and the paper does not claim to introduce a new quantum error-mitigation or compilation algorithm. Its related work includes IBM Quantum Composer and Cirq for circuit construction and visualization, Quantivine for semantic abstraction of large circuits, and geometric representations such as VENUS and Dimensional Circle Notation for quantum states and entanglement. It also discusses noise-learning methods and VACSEN, which already uses coordinated views to examine temporal noise evolution. QVis concentrates on comparing the temporal performance behavior of qubits within a device, relating those comparisons to physical connectivity, and making externally computed circuit-optimization results inspectable. The contribution is the integration and application of visualization and analysis techniques to these tasks, rather than evidence that previous tools offered no coordinated noise analysis. ## Data and coordinated dashboard The hardware examples use characterization data for the 127-qubit IBM Washington processor from 1 January 2022 through 30 April 2023. The source dataset includes state-preparation and measurement error, gate error, gate duration, qubit lifetime $T_1$, and coherence time $T_2$; the paper's demonstrations focus on readout error and daily $T_1$ and $T_2$ records. Higher $T_1$ and $T_2$ values indicate longer retention of quantum information, whereas lower readout error is preferable. These are historical device-characterization observations, not measurements of current IBM hardware. QVis separates its interface into the Qubit Explore and Optimizer tabs. Within Qubit Explore, a metric selector controls coordinated topology, temporal, pairwise-distance, clustering, and metric-distribution views. Figure 2 is an actual dashboard screenshot showing their arrangement: qubit connectivity appears at upper left, the pairwise matrix below it, time-series panels across the upper center, cluster summaries beneath them, and horizontal boxplots in the distribution view at right. The topology represents qubits as numbered circles and their connections as links. Selecting a cluster highlights its members with the cluster's color in the topology and updates the other views; selecting individual qubits in the topology filters the other displays to those qubits. The metric-distribution view is included in the overview, although the paper gives it less methodological detail than the temporal and clustering components. Figure 2 from the paper shows the Qubit Explore interface with selected clusters linked across the views. ## Temporal exploration and visual aggregation The Multi-scale Time Series View combines a focus heatmap, a context heatmap, and a focus line chart, as shown in Figure 3. Time is mapped to horizontal position and the selected performance metric to vertical position. Dragging a selection in the context panel changes the time range displayed in the focus panels; users can subsequently move the entire selection or adjust its endpoints. The context panel retains the wider temporal setting while the focus panels reveal local changes. Because the underlying records are daily, the minimum temporal granularity is one day. The two heatmaps aggregate observations into bins defined jointly by time and metric value. Darker blue indicates that more qubits fall in a bin, and users can adjust the bin count to change the level of aggregation. Hovering over a bin reveals statistical summaries such as its qubit count and median value. This encoding represents the distribution of values across qubits over time, rather than assigning a separate heatmap row to each qubit. Figure 4 compares an overplotted line display of $T_2$ records for all 127 qubits with its aggregated representation, illustrating why binning makes dense patterns easier to inspect. The accompanying line panel retains access to individual temporal trajectories for the focused range. The authors explicitly acknowledge that the aggregated heatmap sacrifices individual qubit values to improve legibility; coordination with the other panels supports movement between distributions and individual traces. ## Clustering, pairwise similarity, and topology QVis applies time-series $k$-means clustering to group qubits with similar temporal behavior in the selected metric. The interface allows users to choose a distance measure, including Euclidean distance or dynamic time warping, and to change the number of clusters. The authors use Euclidean comparisons for their aligned, equal-length daily records to avoid the additional computation associated with dynamic time warping. They select a default of $k=6$ by trial and error, while leaving that setting adjustable for different analytical objectives. This is an application of an established clustering method, not a newly derived clustering algorithm. Figure 5 displays the resulting clusters as small multiple line charts. Each chart maps time to the horizontal axis and metric value to the vertical axis, with semitransparent gray lines for individual qubits and a red summary line representing the cluster barycenter. For the Euclidean case, this summary is the pointwise arithmetic mean of the member series. Colored cluster titles support identification across views, and checkboxes let users select clusters for linked inspection. This makes both persistent differences and short-lived deviations visible without requiring every qubit to be compared in one crowded chart. The Qubit Similarity Distance View in Figure 6 adds a matrix of pairwise distances computed using the selected metric. Both axes index qubits, and each cell represents the distance between a pair of temporal records. Dark blue indicates a small distance and bright yellow a large distance, so prominent rows and columns identify qubits that differ from many others. This is a distance matrix of performance trajectories, not a direct measurement of physical noise correlation or entanglement. In the readout-error example, the authors identify qubits 4, 9, 12, and 109 as having patterns distinct from the other qubits and relate them to clusters 4 and 5 in the dashboard example. The linked topology helps locate those qubits on the processor, but the paper does not establish a physical cause for their anomalous behavior. ## Circuit-optimization interface The Optimizer tab integrates the IBM Qiskit transpiler. Users load a circuit in Quantum Assembly Language (QASM) format and run transpilation and optimization, after which QVis displays structural metrics and the resulting circuit diagrams. The text discusses optimization levels 1 through 3, with higher levels allocating more effort to finding an optimized implementation; Figure 7 also includes a level-0 row in its comparison charts. This describes the paper's Qiskit integration and should not be treated as a claim about the complete configuration of later Qiskit versions. Figure 7 places the original circuit at upper left, horizontal depth bars at upper center, and stacked gate-count bars at upper right. The gate-count chart separates single-qubit gates in purple from multiple-qubit gates in red, making changes in the more resource-intensive multi-qubit operations visible. Expanded circuit diagrams for levels 1, 2, and 3 are stacked below the charts so that developers can inspect structural changes as well as aggregate counts. The illustrated example has shorter depth and gate-count bars at the higher optimization levels. QVis exposes these transpiler outputs for comparison; the paper does not describe a new optimizer or an implemented feedback procedure that automatically feeds the hardware-view clusters into circuit compilation. ## Contributions and evidence The paper's central contribution is a human-directed analysis environment that connects temporal distributions, time-series clustering, pairwise distances, and processor topology, together with an interface for inspecting circuit-optimization outcomes. The Washington data demonstrate how the hardware views reveal groups of similarly behaving qubits and a small set of qubits with distinctive readout-error patterns. The optimization screenshot demonstrates inspection of circuit depth, gate composition, and transformed circuit structure within the same application. These are concrete system demonstrations on characterization data and a circuit example. The paper does not report a controlled user study, a comparative usability evaluation, a broad optimization benchmark, or hardware execution experiments measuring benefits attributable to QVis. Its claims that improved interpretability can support more efficient algorithm design are motivations and intended benefits, rather than demonstrated gains in developer productivity or execution fidelity. Likewise, reductions in displayed circuit depth and gate count do not by themselves establish reduced cloud cost or improved accuracy for an executed workload. ## Limitations and future work The temporal heatmaps trade individual detail for aggregate readability, and their apparent patterns depend on the selected time range and binning. Clustering also depends on the chosen distance metric and cluster count, with no systematic validation of the default cluster count reported. The demonstrations concern a single historical processor dataset, so they provide limited evidence about behavior across devices or about interactive scalability beyond that setting. The paper does not quantify analysis runtime, explain whether detected outliers correspond to independently confirmed hardware faults, or evaluate how users act on the results. These are boundaries of the presented evidence rather than observed failures of the interface. The authors propose extending temporal and multivariate analysis with topological analytics and automated methods for revealing correlations and connections across performance metrics. They also envision a publicly accessible system that dynamically retrieves metrics from multiple quantum devices for ongoing monitoring and benchmarking. Those capabilities are future directions; the demonstrated system already visualizes topology but does not establish that the proposed automated topological analysis or continuously updated multi-device service has been implemented.