By Sundip R. Desai, Lockheed Martin Associate Fellow and Guidance, Navigation, and Controls Engineer at Lockheed Martin Space
The use of data-driven modeling to represent complex systems has become prevalent due to the rise of artificial intelligence and machine learning, which we primarily attribute to the advancements in hardware acceleration, better end-to-end software pipelining, and open-sourced architectures. With these advancements in mind, engineers and scientists can now take massive amounts of data collected from a physical device and create a somewhat usable digital twin in a day. We say “somewhat” because the model is not complete, just a mere reflection of the data organizations provided. Data-driven modeling, or ‘surrogate’ modeling, only captures the structural artifacts of the data that organizations present. The model does not contain innate knowledge, reasoning capability, or perception of the world.
A “World Model” is necessary to develop a reliable digital twin of any system. In this context, a World Model suggests a physical embodiment that captures the nature of the system, manifested as logic, knowledge graphs, or physics equations that promote the surrogate model to extrapolate beyond the capability of what learnings.
In addition, the World Model provides a basis for reasoning, providing judgment, and steering the output logically. Most data-driven models cannot do this as the data must be voluminous and expansive. The data must capture all possible physical states over all dependent variables, incurring a computational and resource penalty. They may still lack all the information and context for an accurate digital twin. A concrete example is a thermistor model on a spacecraft’s instrumentation panel. A simple data-driven model may train temperature readings. However, the model may break due to seasonality effects, Sun-Earth eclipses, or if the spacecraft’s orbit is adjusted. To remediate this effect, organizations should incorporate a World Model to substantiate the data-driven digital twin.
Many methods exist that instantiate a World Model for data-driven digital twins. One such prevalent method is Physics Informed Neural Networks (PINNs). PINNs promote models to conform to closed-form differential equations that govern the physical system. Another method is to capture physical intuition through Neuro-Symbolic AI (NSAI). In a data-driven model, NSAI incorporates logic (i.e., 1st order logic) that embodies rich and constrained physical information. Specifically, NSAI includes prior knowledge and increases the generalizability of data-driven models.
So, let’s return to the thermistor example; what can organizations do to substantiate that model? In addition to collecting sufficient data, organizations can invoke logical constraints or rules to ensure the temperature scales appropriately with attitude and orbital position. Organizations will evaluate the output of the data-driven model across the World Model to provide the most logical output given the conditions, alleviating edge cases mentioned previously and providing a more reliable and explainable digital twin. It is foreseeable that the marriage of the two approaches will be necessary to advance the digital twin landscape.
Become a Digital Twin Consortium Member and take part in the conversation.
The views expressed in this blog are from Sundip R. Desai, a Lockheed Martin Associate Fellow and Guidance, Navigation, and Controls Engineer at Lockheed Martin Space.
About the Author
Sundip R. Desai
Associate Fellow and Guidance, Navigation, and Controls Engineer, Lockheed Martin Space
Sundip’s research includes surrogate modeling of highly complex physical systems, automatic target recognition for vision systems, and anomaly detection. His expertise also includes algorithm development for complex attitude control systems of spacecraft. Sundip earned his Bachelor of Science in Aerospace Engineering at Cal Poly Pomona and his Master of Science in Aerospace Engineering with a focus in Aerospace Controls at the University of Southern California.