Northwestern and Fermilab quantum data helps build a new AI benchmark for quantum calibration with NVIDIA Ising open models
Summarize this article with:
HomeNews and Stories2026Northwestern and Fermilab Quantum Data Helps Build a New AI Benchmark for Quantum Calibration with NVIDIA Ising Open Models Northwestern and Fermilab Quantum Data Helps Build a New AI Benchmark for Quantum Calibration with NVIDIA Ising Open Models April 14, 2026Sara SussmanResearchers at Northwestern, working at Fermilab's underground NEXUS facility, are generating rich, high-dimensional datasets from custom superconducting quantum devices used in quantum computing and sensing. These datasets have helped train NVIDIA Ising, a new family of open models for quantum computing. At the same time, this experimental data is being made globally accessible through the American Science Cloud, creating a shared foundation for AI-driven quantum research.
Bringing Quantum Data to the American Science Cloud Northwestern researchers in the CosmiQ group have begun hosting superconducting qubit data on the American Science Cloud (AmSC) through Fermilab. The first dataset comes from a month-long measurement campaign at the Northwestern EXperimental Underground Site (NEXUS), 107 meters below the surface at Fermilab. This work resulted in a recent Nature Communications paper reporting the first measurement of correlated charge jumps across multiple superconducting qubits in a controlled underground radiation environment. The dataset, which includes measurements of charge jumps, is now globally accessible. Building a New AI Model Benchmark from NEXUS Qubit Data Northwestern and Fermilab’s data has enabled the training of NVIDIA Ising Calibration – a vision language model (VLM) for automating the calibration of quantum processors. Ising Calibration combines a natural-language agent with a VLM that can analyze experimental plots. This is especially useful in quantum experiments where the shape of a 2D measurement often reveals more than numerical summaries alone. The VLM is designed to analyze complete experimental figures and perform direct diagnosis, determining whether an experiment succeeded or whether parameters need adjustment. The NEXUS dataset provides a new test case for NVIDIA Ising. Each scan produces a sinusoidal pattern that shifts discontinuously when a charge jump event occurs. Using this data, the team validated the ability of Ising Calibration to: Identify charge jump events Distinguish clean scans from those containing jump events Flag anomalies that may point to unidentified noise sources The benchmark combines real NEXUS experimental data with synthetic examples to represent a wider range of conditions, including clean scans that were rare in the experimental dataset. For each image, Ising Calibration is tested on six diagnostic tasks: structured description of the figure, classification of the experimental outcome, physical interpretation, assessment of data quality, quantitative extraction of charge jump counts and positions, and a yes/no judgment on whether the qubit environment was stable over the timescale of the scan. Why this matters: These tools could significantly reduce the time it takes to diagnose and tune quantum systems–one of the major bottlenecks in the field. NEXUS experimental lead and Northwestern graduate student Grace Bratrud said, "This will be a great tool for identifying jumps in future datasets and could even enable real-time jump identification, which opens the door to more complex and interesting studies." For example, researchers could monitor charge jumps in one qubit while simultaneously watching for parity switching in another, enabling studies of both charge and quasiparticle dynamics on the same device. Caption: Examples from the NEXUS charge tomography benchmark. Left: a synthetic clean scan with no charge jumps. Center: real NEXUS data with a few charge jumps, visible as sudden shifts in the stripe pattern. Right: real NEXUS data with many charge jumps. The benchmark includes ground truth references constructed from real measurement data and will be released publicly alongside an academic paper, establishing a community standard for testing how well vision models interpret quantum experiment results. AI Model Deployment Infrastructure This collaboration is supported by maturing infrastructure for AI workloads at Fermilab, which Northwestern researchers access through the partnership. NVIDIA Ising is deployed on Fermilab's centralized GPU cluster and registered with the lab's model router, which provides a single endpoint for researchers to access hosted models. Ising Calibration, which powers the NVQCA agent, is deployed through the same pathway. Together, these tools allow researchers to quickly apply advanced AI models to real experimental data, without needing to manage their own GPU resources or model serving. They can point their tools at the router and start working with the models.
Toward Vision Models for Rapid Quantum Diagnostics Looking further ahead, Northwestern and Fermilab are exploring how vision-language models could support more advanced quantum diagnostics for sensing and computing experiments. In many quantum experiments, even partial measurements may contain enough visual structure to determine whether a device is behaving correctly and what follow-up measurements would be most useful. This would make experiments faster, more efficient and more adaptive. This is a promising direction for future work, building on the Northwestern-Fermilab partnership within the SQMS and CosmiQ programs, GPU-accelerated model deployments at Fermilab, and AmSC-hosted datasets. This is the first benchmark in the ongoing collaboration with NVIDIA. Future efforts will incorporate quantum sensing data from Fermilab's QUIET, NEXUS, and LOUD testbeds, all publicly available on the American Science Cloud.