Abstract
Modern data centers and high-performance computing (HPC) clusters demand accurate and flexible digital twins capable of modeling the intricate interplay between workloads, environmental conditions, and physical metrics. Traditional machine learning approaches for HPC telemetry are typically limited to static sets of anonymous sensor variables at fixed positions, optimized for a single task. As a result, these models become obsolete when either the target task or the sensor configuration changes.
In this work, we introduce SeT-Diff, the first foundation model purpose-built for compute-node telemetry and time-series data. Departing from rigid architectures, SeT-Diff employs a diffusion-based generative process conditioned on the semantic description of each sensor, effectively decoupling system dynamics from the specific structure of the dataset.
Our experiments on a real-world supercomputer dataset demonstrate strong performance, achieving a Mean Absolute Error (MAE) of 0.0470 on reconstruction tasks. Moreover, SeT-Diff exhibits zero-shot permutation stability – maintaining high accuracy with negligible degradation even when sensor positions are completely shuffled. A single pre-trained model can seamlessly perform data imputation, forecasting, and virtual sensing tasks. Notably, we report an MAE of 0.033 on thermal inference, validating SeT-Diff as an effective data-driven digital twin for HPC systems.
Methodology
SeT-Diff conditions its generative process on natural-language-style semantic descriptions of each sensor (e.g., "CPU temperature socket 0" or "DRAM power consumption"). By attending to these textual embeddings, the model learns sensor-invariant temporal dynamics, allowing it to generalize to new sensor configurations without retraining. This semantic conditioning is the key innovation that enables cross-task and cross-sensor transferability.
Key Results
| Task | Metric | Value |
|------|--------|-------|
| Reconstruction | Mean Absolute Error | 0.0470 |
| Thermal inference | Mean Absolute Error | 0.033 |
| Permutation stability (zero-shot) | Accuracy degradation | Negligible |
Implications for 2026 and Beyond
As HPC systems continue to grow in scale and heterogeneity – a trend accelerating through 2026 – traditional static telemetry models become increasingly costly to maintain. SeT-Diff's ability to adapt to varying sensor configurations, tasks, and environments in a zero-shot manner positions it as a foundational building block for next-generation autonomous data center management. Its semantic conditioning approach offers a blueprint for building general-purpose AI for operational physics, extending beyond HPC to domains such as smart grids, climate monitoring, and industrial IoT.
Availability
The preprint is available on arXiv (2607.22548) under the subject categories: Artificial Intelligence (cs.AI), Machine Learning (cs.LG), and Performance (cs.PF).
Published: May 11, 2026
via ArXiv AI