Twin Fidelity & Levels

Fidelity is how faithfully and how completely a twin represents its building. It has several dimensions that people often blur together. There is data fidelity - how many points, how frequently sampled, how accurate: a twin reading one main meter every hour is lower-fidelity than one reading hundreds of sub-meters and sensors every few seconds. There is model fidelity - how deeply the twin captures the building's behaviour: a simple set of thresholds is low fidelity; a calibrated physics model of the thermal mass is high. And there is geometric fidelity - from none, through a schematic, to a photorealistic 3D model.

The critical insight is that these are independent, and you do not need them all high. A superb energy twin might have high data fidelity, modest model fidelity and zero geometry. A design-review twin might have gorgeous geometry and almost no live data. Vendors love to push geometric fidelity because it demonstrates well, but it is frequently the least valuable of the three for operations.

Higher fidelity always costs more - more sensors, more integration, more modelling effort, more maintenance - and it adds risk and complexity. So fidelity is not a virtue to maximise; it is a dial to set. The right question is never how realistic can we make it, but how faithful does it need to be to answer the question we are asking.

There is a subtler trap hiding in the model-fidelity dial in particular: a higher-fidelity model is not automatically a more accurate one. A detailed physics model fed with sparse or badly-calibrated data, or never validated against how the building actually behaves, can be confidently wrong - and its very sophistication makes people trust it more, not less. A simple model that has been checked against reality often beats an elaborate one that has not. So fidelity and trustworthiness are not the same thing; adding detail without adding the data and validation to support it buys complexity and false confidence rather than insight. Keep the two ideas separate, and be suspicious of any high-fidelity model whose builders cannot show you how well it matches the real building.

A widely-used way to describe how capable a twin is - as opposed to how detailed - is a maturity ladder with four rungs. Each rung adds a capability to the one below.

1. Descriptive. The twin shows what is happening now: live values, current status, a spatial or dashboard view of state. It answers what is going on? This is the baseline, and honestly it is where a great many valuable twins live - simply seeing the whole building's live state in one place is already worth a lot.

2. Informative (or diagnostic). The twin adds context and explanation: analytics, benchmarking, fault detection and diagnostics, so it can tell you not just that a zone is too warm but why - a stuck damper, a scheduling error. It answers what is wrong and why?

3. Predictive. The twin adds a model that looks forward: simulating or forecasting behaviour to answer what will happen next, and what if? - predicting a component failure, forecasting tomorrow's energy, testing a setpoint change before making it. This needs real modelling (physics-based, data-driven, or both) and good data behind it.

4. Autonomous. The twin closes the loop and acts: within defined limits it makes decisions and writes commands back to the building - optimising control on its own, as in model predictive control. It answers what should the building do, and just do it. This is the frontier, the rarest rung, and the one where security and safety stakes are highest.

Most real building twins today sit on rungs 1 and 2. Rung 3 exists in serious deployments; rung 4 is still uncommon and cautious - and rightly so.

Here is the single most valuable habit in this whole field: match the fidelity and the level to the decision the twin serves, and to the budget - not to what the demo can show. Every rung up the ladder and every notch of fidelity multiplies cost, complexity and maintenance. Climbing higher than your decision needs is how twin projects burn budgets and disappoint sponsors.

Work it the other way round. Start from the decision. If the goal is simply to watch energy and spot obvious waste, a descriptive twin with modest data fidelity and no geometry is exactly right - and cheap. If the goal is to catch failing plant before it breaks, you need the informative rung: rules and FDD, decent data, still no need for 3D or physics. Only if the goal is genuinely to pre-empt - forecast failures, optimise ahead of weather - do you need the predictive rung and the higher-fidelity model it demands. And autonomous control you take on only where the payoff is large, the risk is understood, and the safety and security work has been done.

A worked example: a mid-size office wants to cut its energy bill. A descriptive-plus-informative twin - live sub-metering, benchmarking and fault detection - typically captures most of the achievable savings for a fraction of the cost of a predictive physics twin. Spending ten times as much to add simulation may buy a few extra percent that never pays back. Right-sizing the twin is the value engineering.

Fidelity and level are not one-time choices carved in stone; a well-designed twin grows. The smart pattern is to start low on the ladder and climb only when a proven decision justifies the next rung. Stand up a descriptive twin first - live state, in one place - and let it earn trust and reveal what the building is actually doing. That very quickly surfaces recurring problems, which motivate the informative layer of analytics and FDD. Once you have clean data and understood behaviour, you have the foundation a predictive model needs; without that foundation, prediction is guesswork dressed up in a model.

This staged approach has two big advantages. It spreads cost and risk, so you are never betting a large budget on an unproven twin. And each stage builds the thing the next stage depends on: good data enables good analytics, which enables credible prediction, which is the only safe basis for any autonomy. Trying to leap straight to a predictive or autonomous twin on a building whose data is patchy and whose model has never been validated is the classic overreach - it produces confident-looking outputs that nobody should trust.

So treat the ladder as a roadmap, not a menu. Decide where the building needs to be eventually, but build up to it, letting each rung pay for the next. A modest twin that is trusted and used beats an ambitious one that is impressive and ignored.

Growing a twin also changes how you should judge success. A twin bought as a fixed product is judged on the day it is delivered; a twin grown as a capability is judged on the decisions it improves over years. That reframing matters for how you contract and budget: favour platforms and data models that can climb the ladder later without being torn out, keep the semantic model open rather than locked to one vendor, and set aside operating budget for the analytics and modelling you will add as needs prove themselves. The building that instruments sensibly and starts descriptive, with a clear path upward, almost always ends up further up the ladder - and better trusted there - than the one that tried to buy the summit on day one.