CarbonSync

CarbonSync: Meteorological-Aware Compute Orchestration

CarbonSync is a dynamic workload balancer designed to synchronize frontier model training with global energy cycles. By leveraging Spherical Fourier Neural Operators, we predict renewable capacity factors with high precision, allowing labs to route hyper-scale training jobs to regions where green energy is peaking and compute costs are at their lowest.

What Inspired Us The compute intensity of foundation models has outpaced grid evolution. While the total energy requirement of a trillion-parameter model is fixed, the timing and location of that consumption are variables that can be optimized. CarbonSync moves the compute to the energy rather than straining the grid. The platform transforms the volatility of renewables into a competitive advantage for machine learning teams, turning global weather patterns into a financial and environmental arbitrage tool.

How We Built It We have bridged the gap between global atmospheric physics and data center orchestration through a three-tier technical stack. The system utilizes a Spherical Fourier Neural Operator architecture, trained on the ERA5 re-analysis dataset. Unlike standard CNNs, SFNOs respect the spherical topology of the Earth, eliminating padding artifacts at the poles since they rely primarily on spherical harmonic transforms. Since ERA5 data is coarse and only available at 6h resolutions, we integrated our head with an adaptive FNO. To predict irradiance, we utilized Nvidia's open source AFNO1H model from earth2engine.

In principle, we condition our data by translating atmospheric variables, such as wind speed at 100m and downward solar radiation, into regional capacity factors trained on data from Electricity Maps. We effectively learn the mapping of estimating percentage renewable output given the weather.

Challenges we faced. We spent a lot of time rationalizing which data was okay to use and why. Perhaps the biggest constraint was in actual being able to learn the map between the weather prediction and meaningful capacity factors, while still staying within realistic + reasonable constraints. We spent our greatest amount of time on the weather forecasting and the data processing across these very wide, large datasets -- on which normalization was very tricky since we had to be certainly sure of all downstream implications of the signal we were learning. Particularly interesting was learning to deal with power grid capacities across various lat/lon and being able to estimate meaningfully the capacity factor at this resolution and percentage renewable consumption by data centers.