Define the twin by its update and decision loop
A digital globe displays geospatial information. A model simulates a system. A digital twin should do more: maintain a versioned representation of a real system, update that representation with observations, run constrained scenarios, expose uncertainty, and return evidence to a decision process. For Earth systems, there is no single complete physical twin. Atmosphere, ocean, land, infrastructure, ecosystems, and human behavior are observed at different scales and governed by different models. The architecture is therefore a federation of data and models whose interfaces and validity domains matter as much as graphical output.
The first design artifact should name the decision. Flood evacuation, reservoir operation, crop-risk assessment, transmission planning, wildfire preparedness, and long-term adaptation need different lead times, variables, spatial supports, and loss functions. A high-resolution temperature field may be irrelevant if the decision requires building-level exposure and population vulnerability. Specify actor, action, decision deadline, alternatives, constraints, acceptable false alarms, and the value of improved information. Only then can the team determine which observations, models, and update cadence constitute a useful twin.
Observations are sampled, not complete
Landsat 8 and 9 together add nearly 1,500 scenes per day to the USGS archive. Their offset gives an eight-day combined acquisition opportunity, with 30-metre ground sampling for core multispectral bands. Sentinel-2 provides 10- to 20-metre land bands and a nominal five-day revisit at the equator with two satellites. These figures describe instrument and orbit capability. They are not guaranteed usable cadence: clouds, haze, view geometry, sun angle, outages, quality filtering, and the need for change between clear observations can lengthen the effective interval.
A twin should ingest observation metadata and quality, not only a derived pixel value. Preserve acquisition time, footprint, processing level and version, calibration, cloud and quality masks, geolocation uncertainty, retrieval method, and lineage. In-situ observations remain essential for variables that satellites infer indirectly, for vertical structure, for nighttime or persistent-cloud coverage, and for calibration and validation. Sampling bias should be visible. If monitoring stations cluster in wealthy urban areas or cloud-free regions dominate a product, the twin can look globally complete while uncertainty is distributed inequitably.
Model state and observed state must remain distinguishable
Data assimilation combines prior model state and observations using their respective error structures. The result is an estimate constrained by both, not a direct observation. Reanalysis reconstructs past conditions with a consistent model and historical observations; forecasts propagate an initialized state; climate projections explore conditional futures under emissions and socioeconomic pathways; impact models translate physical variables into hazards or sector outcomes. A twin may connect all four, but its interface should label them. Colour scales that render a modeled 2045 flood depth and a measured 2025 water level identically invite category error.
State updates need a reproducible contract: input version, assimilation window, quality control, bias correction, observation and model error assumptions, rejected data, model code, configuration, initialization, and output timestamp. Ensembles should be retained where they represent plausible state or parameter uncertainty. Averaging can remove the extremes that drive decisions. Model drift and structural error require independent diagnostics, and machine-learning components need evaluation under distribution shift. A fast surrogate is useful only inside the domain in which its error is bounded against the physical model and observations.
Kilometre scale is a scientific advance, not local truth
Destination Earth's Climate Change Adaptation Digital Twin provides a consistent dataset for 1990-2049 at approximately 5 km for atmosphere and land and 5-10 km for ocean and sea ice, with hourly output. Its models and integrated applications support areas including wind energy, wildfire, hydrology, and extreme precipitation. This is a major increase in global granularity, but a 5-km cell still aggregates terrain, coastline, land cover, buildings, drainage, and population. Resolution names grid spacing; it does not establish effective skill at every variable or location.
Downscaling or nesting can add local structure, but information cannot be created without relevant observations and physics. Application teams should test whether higher resolution improves the decision variable, not assume it does. Compare against coarser baselines and independent local data using event-based and distributional metrics. For extremes, evaluate intensity, timing, location, duration, and compound conditions. Bias correction should preserve physical relationships needed by the impact model. When local infrastructure or behavior dominates risk, couple the Earth-system output to audited asset, drainage, exposure, and vulnerability data rather than over-interpreting atmospheric detail.
Compute and data movement are part of scientific design
Destination Earth estimated that one 30-year simulation at 5-km resolution required about 1.5 million GPU-hours or 34 million CPU core-hours during its first phase. The figure is platform- and configuration-specific, but it demonstrates why workflow architecture matters. Moving every native field after a run can be slower and more energy-intensive than computing selected indicators while the model executes. Destination Earth uses one-pass algorithms and a distributed engine to reduce and transform data near the supercomputer, preserving selected high-resolution information without exporting all raw output.
A production twin needs budgets for compute, storage, network transfer, archive retention, reprocessing, and user queries. Every reduction is an analytical choice: deleting temporal detail can prevent later extreme-event analysis, while retaining everything can make the system unusable. Define a tiered archive containing restart state, essential native variables, application-ready indicators, provenance, and reproducible recipes for regenerable products. Monitor energy use and cost per simulation or product. Operational sustainability includes the ability to rerun, audit, and migrate workflows as hardware and software change.
Evaluate decisions, not visual engagement
Validation should proceed through layers. Observation products are checked against traceable reference measurements. Model fields are evaluated across variables, regions, seasons, and extremes using data not assimilated into the same product where possible. Impact models are tested against observed impacts. Decision rules are backtested on historical or synthetic events, including costs of false alarms and missed events. Finally, prospective pilots measure whether the twin changes timing, allocation, design, or outcomes. A more accurate forecast may deliver no value if it arrives after the operational deadline or cannot be interpreted by the responsible institution.
Each result should carry an evidence card: decision and spatial boundary, observation period, model and scenario, resolution and effective scale, update time, validation sample, error metrics, uncertainty representation, known exclusions, and responsible version. Interfaces should permit comparison among models and against observations rather than present one authoritative layer. The most credible twin is not the most photorealistic. It is the one that makes disagreement inspectable, records what was known at decision time, and can show through controlled evaluation whether its information improved a real choice.
Scope and limitations
Spatial resolution, revisit, scene volume, and compute figures describe particular missions and configurations. Effective resolution and usable cadence vary by variable, geography, cloud, geometry, processing, and validation density. Climate simulations are conditional on models and scenarios rather than predictions of a single future, and kilometre-scale output does not remove structural uncertainty. Decision value must be assessed for a named application with local exposure, vulnerability, and institutional data.
References
Source review: 20 August 2026. Quantitative values retain their original definitions, periods, and boundaries.
- 01Climate Change Adaptation Digital Twin
Destination Earth and European Centre for Medium-Range Weather Forecasts · 2026
destine.ecmwf.int ↗ - 02The Fast Development of DestinE's Climate Change Adaptation Digital Twin
Destination Earth and European Centre for Medium-Range Weather Forecasts · 2024
destine.ecmwf.int ↗ - 03The Destination Earth Digital Twin for Climate Change Adaptation
Geoscientific Model Development · 2026
gmd.copernicus.org ↗ - 04Landsat 9
U.S. Geological Survey · 2026
www.usgs.gov ↗ - 05Copernicus Land Services and Sentinel-2
European Space Agency · 2026
www.esa.int ↗ - 06The Space Economy in Figures: Space as a Provider of Critical Data
Organisation for Economic Co-operation and Development · 2023
www.oecd.org ↗