Digital Twin Software Stack
A twin is assembled from ingestion, modeling, assimilation, orchestration, storage, and interface layers, each with distinct requirements.
The stack that runs a twin
A twin is a software system as much as a physics system. It comprises several cooperating layers, and separating them cleanly is what lets the fast, safety-relevant parts stay reliable while the analytics evolve. The layers below describe a representative stack for a plant twin.
The layers
- Ingestion: acquire, timestamp, calibrate, and validate sensor streams at many rates
- Time-series and historian storage: retain raw and derived signals for replay and analysis
- Model services: full-physics codes and fast surrogates, each callable with versioning
- Assimilation engine: run the filters that fuse data with models into a state estimate
- Orchestration: schedule the real-time path separately from batch analytics
- Interface and API: deliver estimates, forecasts, and displays to operators and controllers
Two paths, one system
The real-time path, feeding state estimation and control, must be deterministic and meet hard deadlines. The analytics path, running what-if studies, life tracking, and retraining, tolerates longer times and heavier computation. The stack keeps these physically and logically separate so a heavy analysis never delays a control-relevant estimate.
Reproducibility and provenance
Every output must be traceable: which model version, which data window, which parameter set produced it. This demands disciplined versioning of code, models, and data, and immutable logging of decisions. Without provenance a twin's recommendation cannot be audited, which is unacceptable for a safety-relevant plant.
In the Kronos program
The stack for the Hyperion breeder and the burner is being built around the existing three-dimensional model, extending it from a static representation toward the layered, data-fed system described here. Choices such as open standards and clear interfaces are made now so that when sensors exist, near 2030 for the first breeder, the ingestion and assimilation layers can be connected without redesign. See data pipelines, data standards, and real-time computing.