The Signal
FIU Team Tests Checkpointing for Quantum Neural Networks
Florida International University researchers are developing checkpoints to help quantum workflows recover after interruptions. Their quantum-neural-network framework requires 186.6 KB of storage for a 100-qubit model, according to the team’
Quantum.world Team · Written with AI · fact-checked ·
“Scientists working on everything from new drugs to hurricanes would benefit from saving their progress because today’s quantum computers are error-prone and can’t save. That’s what makes quantum checkpointing so exciting, the problem stretches from quantum theory to hardware limitations and finding a solution will impact science and society.”
Janki Bhimani, director of FIU’s Data Management Research Laboratory
Florida International University computer scientist Janki Bhimani and her team are developing ways to restart quantum workflows after interruptions, addressing a weakness that can waste scarce access to quantum processors. FIU said on October 27 that the researchers have tested their work across multiple IBM quantum computers. The effort focuses on preserving progress in quantum computing tasks without trying to save an unknown quantum state as if it were ordinary computer memory.
Why quantum jobs cannot save a state snapshot
Classical computers can periodically copy program data to a checkpoint and reload it after a crash. Quantum computers cannot directly use that method for an arbitrary unknown quantum state: quantum mechanics prevents exact copying, and measuring a state can change it. A failed or interrupted run can therefore force researchers to repeat work, even when the workflow has already completed useful stages.
FIU’s description frames checkpointing as a systems problem, not a way to preserve a qubit’s full physical state. The goal is to help a workflow resume after interruption, particularly in hybrid applications that combine quantum operations with classical computation. That distinction matters: a restart strategy can preserve useful computational progress without claiming to restore the exact state that was lost.
A 186.6-kilobyte checkpoint for a 100-qubit model
The university’s Data Management Research Laboratory describes a quantum-neural-network checkpointing framework developed by Christopher Kverne, Mayur Akewar, Yuqian Huo, Tirthak Patel and Bhimani. The lab reports that its framework uses 186.6 kilobytes of storage for a 100-qubit quantum neural network and incurs “negligible overhead.” It also reports that the same model can behave differently across quantum devices, making hardware-specific information and shot counts relevant to a checkpoint.
The team’s paper, Quantum Neural Networks Need Checkpointing, appeared at the 2025 ACM Workshop on Hot Topics in Storage and File Systems. The lab says its experiments used several quantum devices; FIU’s October announcement identifies the systems as IBM quantum computers but does not name individual processors. Neither FIU’s announcement nor the lab summary gives a numerical overhead figure, so the reported storage amount is the clearest quantified measure of the framework’s footprint.
Recovery depends on more than stored data
Hardware dependence complicates the familiar idea of saving a checkpoint and restoring it anywhere. If a quantum neural network’s behavior varies between devices, a useful record may need to capture details about the hardware and how the computation was run, not merely its model data. That makes checkpointing part of the workflow’s design and its record-keeping, rather than a generic backup step added after the fact.
FIU points to drug discovery and climate modeling as possible areas where researchers could benefit from preserving long-running computational work. Those are prospective applications, not results demonstrated by the checkpointing research described in the announcement. The concrete result so far is narrower: a small-storage checkpointing framework for quantum neural networks, tested across real devices, with the team’s reported storage requirement of 186.6 kilobytes for a 100-qubit model.