Terra Daily — July 15, 2026

Research Worth Reading

• Contraction Certification from Streaming Data: Wasserstein Robustness and Compositional Stability for Interconnected Nonlinear System — Presents streaming contraction certificates that determine in real time whether observed data suffices to certify a safe control action under shifting disturbance distributions in interconnected nonlinear systems. The Wasserstein-robustness and compositional stability framing transfers to safety-critical control for grid or autonomous environmental monitoring where disturbances are non-stationary.

• WULPUS PRO: Multi-mode Ultra-Low-Power Wearable Ultrasound and Array Imaging with CMUT Support — A fully wearable ultra-low-power ultrasound platform supporting multi-channel CMUT array imaging, moving beyond the shallow, low-channel A-mode sensing of prior wearable platforms. The low-power front-end and array processing are directly reusable for constrained-power environmental sensing nodes and remote monitoring hardware.

• Generating Physically Plausible Parachute Dynamics with Deep Generative Modeling — Uses deep generative modeling to learn physically plausible parachute and entry-vehicle dynamics that resist traditional system identification. The pattern—learning dynamics from limited, noisy data while respecting physical structure—transfers to modeling poorly-characterized physical systems in climate engineering.

• Multi-Fidelity Uncertainty Propagation with Model Adaptation to Local Cislunar Dynamics — Propagates uncertainty for cislunar space objects using multi-fidelity models with local adaptation for space situational awareness. The multi-fidelity surrogate approach applies to any sparse-data physical simulation problem, including satellite operations for Earth observation.

• Dynamically Feasible Planning and Control in Complex Environments: a Scalable Systematic Approach — Builds safe sets for linear discrete-time systems under non-convex constraints represented as a union of polytopes, then implements a reference governor for safe reference tracking. The constrained-control formulation maps to energy system problems like battery dispatch or EV charging where operational limits are non-convex.

• A Hierarchical Semi-Markov Load Model for AI Data Centers Coupling Job Scheduling with Bulk-Synchronous-Parallel Power Dynamics — A hierarchical semi-Markov model couples AI training job scheduling with bulk-synchronous-parallel power dynamics across compute, sync, and checkpoint phases. Understanding and shaping this load is a concrete systems lever for grid-interactive data centers and reducing the carbon intensity of compute demand.

Today’s Synthesis

The hierarchical semi-Markov load model for AI data centers shows that training-job scheduling produces non-stationary, phase-dependent power draws that a grid operator sees as a shifting disturbance distribution. Pair that load characterization with the reference governor built on safe sets for non-convex constraints : representing data-center operational limits (peak demand, ramp rates, backup battery state) as a union of polytopes gives a tractable way to track a grid-friendly power reference without violating hardware limits. The missing piece is certification under distribution shift, which the streaming contraction certificates address directly—they determine in real time whether observed load data still warrants a safe control action as job mix changes. An engineer building grid-interactive compute can prototype this stack today: use the semi-Markov model to simulate phase-aware load, the polytope safe sets for dispatch bounds, and the Wasserstein-robust contraction test to gate adjustments. The systems skill of writing a reference governor transfers cleanly from classical control to carbon-aware compute.