Using Microsoft Fabric to create digital twins

Digital twins are an important part of modern control systems. They let you create a model of a physical system that serves as a tool for simulation and prediction. It models the inputs and outputs of a system, using real-time data to keep it at the same state as the “real world.”

Building a digital twin isn’t easy, especially when you’re modeling complex systems at scale. The first twins were simple, with a handful of inputs and outputs. Now we’re modeling complete industrial processes or entire energy-generation systems. A tool that used to deal with a single device now must support some of the largest infrastructures on the planet.

Modeling large-scale, complex systems

Modeling complex systems is hard. We’re dealing with physical objects and dynamic processes. The digital twin of a steel blast furnace will need to include the supply chain that delivers the coke that fires the furnace so that we control how much we need to burn without letting the residue in the furnace cool and turn to glass, a failure that requires months of work and millions of dollars to get back to production.

These large models need a lot of data, not only to create the digital twin but also to stream the data that keeps them up to date. Once we have that data we can use our digital twins as part of a control system or in conjunction with machine learning to identify possible error conditions and predict failures well in advance. This allows for preventative maintenance well in advance of a possible breakdown.

To deliver large-scale digital twins, we need a large time-series storage environment to deliver the necessary fidelity. It needs to handle many different data types from many different sources and still support a common set of query interfaces to run models at cloud scales. The system needs to be able to work with possibly fragmented data and within regulatory frameworks that may well be more complex than the systems they’re modeling.

Introducing Fabric’s digital twin builder

At Build 2025, Microsoft announced new digital twin capabilities for its Fabric data platform. It provides analytics on top of a mix of data types, including real-time and time-series data, working in large-scale data lakes.

Digital twin builder is part of Fabric’s real-time intelligence tool. Like much of Fabric’s analytic services, it’s designed to be a low-code development platform where stakeholders and subject experts can build the tools they need in collaboration with traditional developers. At the heart of this platform is a set of tools that lets users define the key concepts needed for a digital twin model as an ontology.

This ontology maps to dashboards, which can show visualizations of what is happening to a digital twin, delivering information to workers who need to make quick decisions. With modern digital twins focused on entire environments, this can help manage situations such as weather conditions in offshore wind farms, controlling blade speeds to avoid overloads and protecting both the hardware and the wildlife around the turbines.

The new tool sits inside Fabric, on top of its real-time intelligence module. This gives it access to the built-in data connectors, and the data can be accessed by other applications that have access to Fabric’s OneLake store.

Building a digital twin with Fabric

Like most Fabric applications, the digital twin builder needs to have all its data stored in a lakehouse, with Fabric tools defining the ontology that maps data to the real-world systems and the processes your digital twin will model. Having experts on hand is important as you build the vocabulary that Fabric uses to pull together many different data sources and streams. This mapping process is key to defining the essential entities your model uses, bringing them together with queries to define semantic relationships.

Entities are specific machines, processes, inputs, and outputs, as well as the people and systems that sit around them. These are linked to the data in Fabric associated with them in a semantic hierarchy that lets you wrap different entities together.

Once you have built this map of your data, you’re able to use other Fabric tools to explore and analyze the data, linking it to the systems and the processes you’re modeling. This can be extended further, out into AI tools for more insights, for example, finding outliers in your data or providing visualizations. APIs let you use that data with code, linking it to digital twins of your control systems and models of physical hardware. Visualizations can be built in familiar tools, such as Power BI, or data can be delivered.

Having this data allows you to work with machine learning systems for training and testing. You can start to build models that predict failures or provide insights into product quality based on inputs and settings, giving you tools to optimize your operations.

Getting started with this new service means enabling it in the Fabric admin portal, as well as installing the latest Power BI on user desktops. It’s important to note that you can’t go into building a digital twin blind; you need to have done a lot of work to identify and understand the data sources you’re going to use and how they relate to the specific system you’re going to model.

It’s a good idea to have this as a diagram that shows the key elements of the system, how they are linked, and where the various data sources and control points are. This will allow you to quickly see how controls affect the information you have: For example, how does adjusting a valve in a pipe affect flow rates at various points throughout a chemical process plant? Diagrams like this can even be used as the basis of a Power BI dashboard.

Having this diagram helps define the ontology you will use to classify data in Fabric. Data will come from multiple sources and in multiple formats, which is why Microsoft uses Fabric—there’s no need to apply expensive and complex ETL processes to normalize data, as it can be stored in its own native format with queries working across the resulting data lake. You can mix time-series operational data from internet of things systems, ERP data from tools like SAP, details of the equipment used from your existing hardware inventory, as well as many other associated data. It all wraps around the process you’re modeling as a digital twin.

Working in the semantic canvas

At the heart of this process is a new tool, the semantic canvas. This is where you create and manage the entities in your ontology, adding relationships to other entities and mapping data to them. The ontology model used by Fabric is at heart hierarchical, allowing you to group entities into namespaces, adding types and having multiple instances. For example, if you create an entity for a specific sensor type, you can have instances for each implementation of that sensor in your physical plant. Data is loaded into the Fabric lakehouse for your digital twin and then mapped to an entity instance in the semantic canvas. The digital twin tool provides different mappings for different types of data, including controls for how data is processed.

Once you have built your entities and mappings, you can start to use the digital twin, first by using Power BI to view reports and then by including entities in a Fabric real-time dashboard. Other tools generate alerts based on data. The ontology data can be used with machine learning predictive models to add context to raw sensor data, either by using Azure’s AutoML tools to automatically choose and tune a model or by using your own custom models.

We have a lot of data in our businesses that can be used alongside sensor data from the industrial internet of things. Having a data lake platform allows us to bring in that data at scale and build much more complex digital twins more easily. The more data we have, at a higher fidelity, the better the model and the more accurate the control. We can operate systems more efficiently and safely without needing significant software development resources.

Building digital twins on Fabric gives us that scale with a low-code development model. By building on Azure, we have the necessary processing power to support both real-time analytics and predictive machine learning. Knowing what’s happening now in an industrial process allows us to link it to the larger-scale business processes it is part of, while being able to choose when to maintain it for the least possible downtime or know when to adjust materials and equipment to meet changes in demand.