A machine-learning model reconstructed the temperature, pressure and velocity of a simulated liquid-metal flow from sparse temperature readings, producing a new estimate in less than a second after training. The result suggests that a fast digital monitor may be feasible for this type of computational test, although the strongest performance was not uniform across magnetic-field conditions.
The work is an arXiv preprint, version 1, dated 28 August 2026.
From a few readings to a full flow field
The framework couples principal component analysis, a way to represent complex fields with a smaller set of patterns, with a shallow recurrent decoder called SHRED. An LSTM recurrent network first processes the time history of the sensor readings, and a decoder then maps that information back into the full physical space.
The benchmark was a three-dimensional square-cross-section channel measuring 0.3 metres long and 0.03 metres high, with two transverse cylinders 0.01 metres in diameter. Its mesh contained 196,200 cells. The reference trajectories came from the transient OpenFOAM solver magnetoHDFoam, using the PIMPLE algorithm.
Each simulated run lasted 3 seconds and was sampled every 0.025 seconds, giving 150 snapshots for each magnetic-field scenario. The model received only temperature measurements; pressure and velocity were inferred indirectly. The temperature readings were given centred Gaussian noise with a standard deviation of 0.02, and four random sensor placements were used for each configuration.
Accuracy held across changing field angles
In one test, the researchers varied the magnetic-field inclination across 17 simulations, covering angles from 5 to 30 degrees while keeping the field magnitude at 0.05 tesla.
Across that angle analysis, the mean relative error stayed below 3.3% for both random and DEIM sensor placement. DEIM reached saturation with three training angles, while the difference between the two placement strategies remained below 0.25%.
The study measured accuracy with a mean relative error based on the L2 norm, a measure of the size of a field difference. In plain terms, the calculation compared the reconstructed field with the full-order-model reference field, divided the difference by the reference field size, and averaged the result over test times and parameter scenarios.
More difficult magnetic fields exposed weaknesses
A second ensemble contained 21 high-fidelity simulations at a fixed inclination angle of 5 degrees, spanning different flow regimes as the magnetic-field intensity changed.
Changing the share of cases used for training from 20% to 70% made little difference to the mean relative error for most fields. The stated exception was pressure at 0.3 tesla, where the error remained at or below 6.5%.
The broader test varied both parameters at once, using 14 magnetic-field values and four inclination angles for 56 simulations. The set contained 40 training cases, eight validation cases and eight testing cases.
At a weaker field of 0.075 tesla and angles from 5 to 30 degrees, the mean relative error stayed below 5% for temperature, pressure and velocity. Those errors were only slightly above the lower bound imposed by the study’s SVD rank truncation.
The harder cases were at higher field strengths. Most mean relative errors were below 5%, but pressure errors were around 8% at inclination angles of 21.7 and 30 degrees, while the velocity error approached 10% at 5 degrees.
The model also recovered a changing flow pattern
For a 0.3-tesla field inclined at 30 degrees, the simulation showed laminarization, meaning a more orderly flow pattern, along with an asymmetric side layer confined to the upper part of the domain. The researchers reported that SHRED captured these features, and that the spatial patterns of reconstruction residuals and ensemble uncertainty were similar.
A fast computational result, with clear boundaries
Each configuration took about 3 minutes to train, while estimating the state at test time took less than 1 second. A full-order transient simulation took more than 6 hours, creating the large timing gap that motivates the surrogate approach.
The findings remain tied to the simulated benchmark, its selected magnetic-field ranges, its full-order solver and its synthetic measurement model. The fixed Gaussian noise and simulated sensors limit what can be concluded about real measurements. The reported ensemble spread was used as an empirical error indicator, without formal probabilistic calibration or coverage analysis.
The study supports computational feasibility for monitoring a modeled three-dimensional magnetohydrodynamic flow, with especially low errors in the weaker-field double-parameter test. It does not by itself establish performance in a physical facility or show that the method improves reactor control, safety or heat extraction.
Paper data and sources
Original title: Real-Time Monitoring of MHD Liquid Metal Flows with Shallow Recurrent Decoders
Authors: Claudio Scardino, Stefano Riva, Carolina Introini et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text