AIBRIEF
A Bologna and Italy's national supercomputing consortium (CINECA) preprint calls its single pretrained network a 'foundation model' for sensors, conditioning on each sensor's meaning to handle filling in missing readings, forecasting, and thermal
Modern data centers and supercomputers run on a steady stream of temperature, power, and workload sensors, and most of the AI used to watch them is rebuilt whenever a sensor changes or a new task is added. A new arXiv preprint from University of Bologna and CINECA researchers describes one general-purpose model that aims to handle all of it.
The system, called SeT-Diff, is a diffusion model that conditions on a short text description of each sensor, its name, its units, and what it measures, processed through pre-trained textual embeddings. That semantic conditioning is the trick that lets the same pretrained network handle three jobs on one supercomputer: filling in missing readings, forecasting future values, and inferring temperature at unmeasured locations. The authors call this decoupling system dynamics from the structure of the dataset.
The preprint reports a mean absolute error of 0.0470 on reconstruction and 0.033 on thermal inference, and claims the model stays accurate when sensors are reshuffled. It is posted as arXiv:2607.22548v1 on 11 May 2026 by Giovanni B. Esposito, Francesco Antici, and Andrea Bartolini at the University of Bologna, with Daniele Cesarini at CINECA.
It is a preprint, not a peer-reviewed paper, and the abstract does not name the comparison baselines for those error figures or the supercomputer whose data the model was trained on. A public dataset and a named baseline comparison are the natural next tests.