Myth: Condition Monitoring and the Digital Twin Are AI

This article comes as a result of attending a number of AI panels in the Turbomachinery world. It is annoying to go to an AI panel, and then all the panelists talk about the digital twin and condition monitoring. It may well be that I went to the wrong panels, because certainly the CFD people discuss more genuine uses of AI.

So, to clear the discussion, first let's define what AI is. I let AI speak for itself (Google AI, accessed 2/12/26): "AI is a field of computer science focused on creating smart machines that can perform tasks that typically require human intelligence, like learning, reasoning, and problem-solving". We can distinguish machine learning, deep learning, natural language processing and computer vision. The current state of affairs is Artificial Narrow Intelligence (ANI), designed to perform a single specific task. It combines data in an algorithm to make predictions within pre-defined parameters. Gemini, ChatGPT and other large language models (LLM) fall in that group. It requires training data to perform. A future step would be Artificial General Intelligence (AGI) for a broad range of tasks, would use human-like reason to learn, adapt and improve, with the possibility to eventually create an Artificial Super Intelligence (ASI), a system that would be self-aware, and would operate beyond human control. All types of AI depend on the quality and quantity of their training data, and results may exhibit inherent biases of the training data. Of special interest are so called AI agents that can take environmental data input (for example sensor data), use existing software models (e.g., a digital twin or its cousin, the performance map), and use this to plan, reason, act and learn or adapt.

Digital twins are physics-based models of a machine, that replicate the performance (in the widest sense) of the machine based on a set of input data. A simple form is the performance map of a compressor, that allows us to predict compressor speed or compressor efficiency if we enter gas composition, flow and process head requirements. These models can be extended to the complete systems, such as a gas turbine driven train with coolers, valves and system behavior. The models can be adaptive, that is parameters in the model can be changed to make it match input data with actual measured machine data. We discussed Digital twins in a 2022 MythBusters article. Applications in the turbomachinery space include analysis of operational data (vibrations, temperature, pressure etc.) to predict a likely failure of a component or system. Digital twins are also used for performance optimization, for example by using data from an entire fleet of assets. This can be used to provide operational guidance to the operator, or to improve the design of future products.

Other digital twins can be generated by neural networks, that learn from large sets of data without the use of physics-based models to correctly match the output data for a set of input data. Issues include the quality and range of training data, as well as the fundamental issue that correlations do not always imply causation. Just because two variables appear to move together, it does not mean that a change in one variable directly causes the change in another. The bigger the training data set, the more statistically significant correlations can be found, and many of them may just be coincidental: Ice cream sales and the rise in sunburnt skin are correlated, but ice cream does not cause sunburn. It's the hot weather that causes both.

Either model can be used for diagnostics, including to some extent the identification of faults in the system. Going back to our simple compressor map, a system may be able to identify fouling, changes in gas composition, or balance piston leakages from differences between model prediction and actual data. Similarly, it may be able to identify sensor faults. Coupling such a model with rotodynamic models then allows insights in the mechanical health of the machine. Simulating an entire system, for example an LNG plant or a natural gas pipeline, consisting of several compressors and drivers allows the operator to optimize the operation of the plant to reach a pre-defined goal, for example minimal energy consumption.

In the design space, AI approaches can for example be used to improve CFD models without the requirements of computationally extremely expensive simulations. The issue for CFD models has always been the modeling of turbulence. AI allows more advanced turbulence models that can be learned from data. In the modeling process, a high-fidelity database must be established using highly accurate experiments or high-resolution simulations (DNS, LES). Such a database must cover sufficient elementary flows such as separated flows, attached boundary layers, or free shear flows. Finally, models can be established by machine learning.

The systems described are deterministic, and in case of malfunction of the simulation tools, errors can be traced and identified relatively easily.

Autonomous decision making with AI, on the other hand, may lead to non-deterministic solutions to problems. These may be solutions that come as a surprise to the observer, either in a positive or a negative sense. Once the solution is revealed, it is often impossible to reconstruct the decision path of the system. It may be a solution that violates goals that the system owner considers self-evident. The AI system does not necessarily know 'self-evident rules' if it was not trained to recognize them.

The current condition monitoring systems and digital twins are not AI. That doesn't mean that AI cannot be used for condition monitoring purposes, for example to improve and adapt performance models, and to identify problems. Digital Twins then can be the tools for AI agents. AI agents are software systems that can sense their environment, analyze data, make decisions, and act to achieve goals, without constant human input. Unlike conventional software, which follows fixed rules, AI agents adapt based on the information they gather and learn from experience. The AI agent may show reasoning, planning, and memory and have a level of autonomy to make decisions, learn, and adapt.