Research Worth Reading

FESOM2-JAX v1.0: a differentiable shadow of the ocean-sea-ice model FESOM2, cast onto GPUs — A full Python/JAX reimplementation of the FESOM2 ocean-sea-ice model that scales from CPUs to 256 GPUs and is end-to-end differentiable. Enables gradient-based calibration, data assimilation, and direct coupling with ML components for climate modeling. Open-source and designed as a ‘code shadow’ that mirrors the Fortran reference.

Physically Consistent SINDy (Sparse Identification of Nonlinear Dynamics) for Microgrid Identification and Real-Time Frequency Control — Physics-guided framework (PC-SINDYc) for identifying microgrid frequency dynamics from data using sparse regression with total least squares and RANSAC. Discovers interpretable, physically consistent governing equations enabling real-time frequency control of inverter-based microgrids with distributed energy resources.

Critical Inertia Estimation for the Three U.S. Interconnections — Simulation-based assessment of minimum system inertia required to prevent under-frequency load shedding across U.S. interconnections as inverter-based resources displace synchronous generators. Quantifies inertia thresholds using detailed dynamic models with high renewable penetration — directly actionable for grid planners.

Smoothing the Ramp, Not the Peak: Scheduling-Induced Power Dynamics of LLM Inference and Their Grid-Scale Consequences — Analyzes real GPU power traces from LLM inference serving to show that chunked prefill scheduling creates controllable power ramps affecting grid stability. Demonstrates how scheduling knobs can shift load profiles to smooth ramps rather than reduce peaks, providing a software-level lever for grid integration.

Technology & Innovation

Q&A: Does the world need ‘carbon capture and storage’ to reach net-zero? — Technical deep-dive on CCS deployment requirements for net-zero pathways, covering capture technologies (post-combustion, pre-combustion, direct air capture), storage capacity estimates, energy penalties, and cost curves. Includes IPCC scenario analysis showing CCS necessity for hard-to-abate sectors.

How to bring a geothermal well back from the dead — Geothermal company Zanskar acquired a failing New Mexico power plant with rapidly degrading shallow wells and applied advanced reservoir modeling and stimulation techniques to restore production. Case study in how data-driven well diagnostics and targeted interventions can revive stranded geothermal assets.

Antora snags $550M for heat batteries to run data centers and factories — Antora Energy’s thermal battery system uses solid carbon blocks to store renewable electricity as heat at >1,500°C, then delivers continuous high-temperature heat or electricity via thermophotovoltaic conversion. Targets industrial process heat (steel, cement, chemicals) and data center loads with a solid-state, no-moving-parts architecture.

Today’s Synthesis

The grid is becoming a software problem. Smoothing the Ramp shows that LLM inference scheduling creates controllable power ramps — a software knob that directly affects grid stability. Critical Inertia Estimation quantifies the minimum inertia thresholds across U.S. interconnections as synchronous generators retire, giving you concrete stability boundaries to design against. PC-SINDYc then provides a physics-guided system identification framework that discovers interpretable governing equations for microgrid frequency dynamics, enabling real-time model-based control of inverter-based resources. Together, these three papers map a clear engineering stack: you can measure the grid’s dynamic limits, identify the system’s physics from data, and build scheduling or control loops that actively shape load ramps to stay within stability margins. For a software or ML engineer, this means grid-forming inverter control, scheduling-aware load shaping for data centers, and physics-informed surrogate models for real-time frequency regulation are all tractable entry points — no power systems PhD required, just solid control theory, optimization, and differentiable programming skills.