AI agent designs and tests MRI sequences in a virtual research workflow

MRI images depend not just on the scanner, but on the instructions it executes. These pulse sequences determine when radiofrequency pulses excite tissue, how magnetic gradients encode position and when signals are collected. Changing that timing can alter tissue contrast, scanning speed and measurements of tissue properties.

SmartBloch-Agent aims to automate the research process connecting an imaging goal to a tested sequence. Jiakai Yang and colleagues at Xiamen University and the University of California, Irvine describe the framework in an arXiv preprint submitted on November 6, 2026. The supplied version is not a peer-reviewed publication, and its demonstrations use numerical simulations rather than actual scanner acquisitions.

An AI coordinator, not an unchecked physicist

The framework gives distinct jobs to different components. A language-model agent, implemented using DeepSeek V4 Flash, interprets objectives, proposes actions and coordinates tools. Sequence-building software constructs explicit events, numerical optimizers search parameter choices, and the team's GPU-accelerated SmartBloch engine calculates magnetization evolution using the Bloch equations.

Simulated signals can then be reconstructed into images and evaluated against task-specific criteria. Failure reports and measured outcomes return to the agent to guide revisions.

Crucially, the agent does not get to declare success by itself. Executable checks enforce timing and parameter bounds, while separate scientific checks assess eligibility. A candidate can be executable yet scientifically unacceptable, or score highly while failing a safeguard. Selected outputs undergo a separate computational replay, although this generally uses the same simulation backend rather than an independent physical experiment.

Designing contrast and acquisition order

One demonstration used fast spin echo imaging of a digital brain phantom. The objective required cerebrospinal fluid to appear brighter than gray matter, and gray matter brighter than white matter, while respecting signal-retention and scan-time limits.

Across three successive candidates, the agent changed repetition time, echo spacing and acquisition organization. Simulated scan duration fell from 160 to 96 seconds, while both specified tissue-contrast measures increased.

A subsequent comparison showed why timing alone is insufficient. Changing which spatial-encoding lines were assigned to particular echoes reduced the fraction of reconstructed image energy outside the phantom's expanded tissue boundary from 1.40180% to 0.04083%. That improvement came with a trade-off: fluid-to-gray-matter contrast increased, but gray-to-white-matter contrast decreased.

Across another 12 planned design tasks covering contrast, scan time, radiofrequency resources and spatial leakage, all produced eligible designs that passed replay. However, these required 15 launch attempts, including interruptions associated with GPU occupancy. The reported automation therefore does not mean every launch proceeded without environment repair or resource recovery.

Better tissue estimates—with important qualifications

A second major experiment optimized excitation flip angles in a five-echo sequence called SSFP-MOLED. The agent worked alongside Optuna/NSGA-II numerical search, with fixed acquisition timing.

Of 191 fully evaluated, deduplicated candidates, 58 met the original multiobjective eligibility rules. The selected language-model proposal improved five of six normalized uncertainty-related measures, with a slight deterioration in the sixth. These measures estimate local signal sensitivity under specified modeling and noise assumptions; they are not direct measurements of clinical accuracy.

The researchers then selected a different, numerically generated candidate for estimating T2, a tissue relaxation property. Its normalized T2 uncertainty proxy was 45.33% lower than the reference. Importantly, it passed absolute execution checks but failed some original reference-relative safeguards, including those concerning field sensitivity, echo balance, signal and robustness. It was selected under a narrower T2-focused criterion, not as the original all-round winner.

To test downstream performance, each sequence generated 512 matched synthetic samples. Separate reconstruction networks used the same architecture, training budget and selection rules. Across 64 paired held-out samples, mean absolute T2 error fell from 5.690 to 5.196 milliseconds, an 8.69% reduction. Error improved in 60 samples and worsened in four.

The result reflects each acquisition paired with its trained reconstructor. It also involved a resource trade-off: the optimized sequence's sum of squared flip angles, an RF-energy proxy, was approximately 4.124 times the reference value.

What remains unproven

No additional measurement noise was introduced in reconstruction training or testing, and the synthetic samples were not established as independent subjects. These results cannot demonstrate patient-level performance or robustness to real acquisition noise.

The simulator met prescribed tolerances in tested single-voxel comparisons against an independent mathematical reference, but that does not validate every sequence or imaging condition. Reported workflow savings also compare recorded agent execution with estimated coordination budgets, not separately timed human studies.

The work supports automated coordination of constrained MRI simulation research. Moving its output packages onto scanners still requires compatible execution software and on-site hardware safety checks; clinical utility remains untested.