A large language model (LLM) is a generative AI model used to produce language-based responses. In a digital twin, it can provide a natural-language interface to system state, generate simulation scenarios, or help create synthetic data when real-world data is scarce. It is not a substitute for the twin’s measurements and asset knowledge. section 1.2.3 section 5.2.3
Give an LLM relevant real-time and historical twin data before asking equipment-specific questions. An answer about a pump’s optimal pressure may otherwise be generic or wrong. Asset details, maintenance records, engineering specifications, and sensor readings can ground a response in the equipment being discussed and support diagnostic analysis. section 5.2.3
From readings to a diagnostic report
An LLM can examine turbine vibration readings alongside operating hours and a maintenance schedule, then generate a report about a possible bearing failure. The readings and records provide the context for the analysis; the generated recommendation still needs appropriate scrutiny. section 5.2.3 section 8.4.4
An LLM-driven agent can reason across changing conditions when fixed rules become difficult to maintain. Keep physical control behind deterministic controllers and safety guardrails: the agent may adjust a temperature setpoint rather than switch an HVAC compressor directly. LLM outputs can vary for the same input and can confidently give unsafe instructions. section 8.4.4
Check whether that flexibility pays for itself. Managed LLM services charge by token, so frequent decisions can cost more than the savings they produce. For a simple HVAC decision, a low-cost rule may be preferable to an agent running every 15 minutes. section 8.5.3
