Preprint Details Data Schema for Multi-Modal Fusion Sensors Spanning Five Orders of Magnitude in Sampling Rate

The work characterizes input complexity across point measurements, spectrograms, and images with nonstationary physics, then quantifies trade-offs between temporal context length and frequency resolution needed to preserve fluctuation information. Analysis draws on real fusion diagnostics such as those at DIII-D and JET, where similar rate mismatches have historically forced manual downsampling that discards high-frequency MHD activity. By treating the problem as a representation rather than a collection task, the authors surface requirements that current scientific data lakes ignore.

Existing literature on scientific foundation models, including the 2024 Nature Machine Intelligence review of physics-informed transformers and the 2025 arXiv survey on multi-modal plasma control, focuses on architecture scale while assuming clean, aligned datasets. Chen's case study reveals that assumption fails for fusion: sensor heterogeneity exceeds ImageNet-style benchmarks by orders of magnitude. The missed implication is that data organization itself becomes the primary bottleneck, not model size.

Adoption of this schema could enable cross-device transfer learning between tokamaks and stellarators, a step current single-lab datasets cannot support. Next milestones include public release of the formatted dataset and integration tests within existing real-time plasma control loops at one of the major facilities within 12 months.