Generalized graph foundation models as versatile data-driven digital twins for complex technological systems.

Digital twins are comprised of computational models that mimic the 'as built' characteristics of devices, systems, and networks of systems whose performance in the real world warrants quantitative and critical assessment.The literature on constructing digital twins is historically focused around task-specific, physics-based models that seek to understand the device from first principles, thereby constructing an idealized digital twin, based on the known physics, of the system.However, these "physics-based digital twins" (pbDT) can be quite difficult to construct, as the level of detail required to accurately model the complex physics of many devices is often missing or expensive to obtain, especially for systems with widespread deployment.Additionally, pbDTs generally assume the device is working as intended; in practice, many systems experience some form of performance degradation, or derating, that causes them to operate off-specification, in manners such that the basic physics is undetermined.As it is often the goal of a digital twin model to quantify these departures from the idealized system, it is quite difficult to separate assumptions from the expected model output.In contrast, data-driven digital twins (ddDT) seek to model the system as it actually is based on real observations and datastreams arising from the device in question.ddDTs enable agility in responses because they learn system dynamics directly from data at operational timescales. As a result, complex physical phenomena of a physical system or process, which are often difficult to explicitly model, can be captured since their combined effects are implicitly reflected in the measured datastreams of sensors and control signals. Additionally, ddDTs generally utilize a flexible model architecture (typically an artificial neural network) to avoid injecting implicit bias into the system. This flexibility also lends itself to another advantage: modularity, that a single ddDT model architecture can be used to train ddDTs for multiple, quite different systems, and to answer multiple different questions about the real-world system.With a ddDT, it is possible to train the model in a self-supervised manner via a reconstruction objective to obtain a trained "encoder" module.This trained encoder module can then be used as a Foundation Model (FM) for the system to answer different task-specific questions without necessitating the training of a new, task-specific, pbDT or training another ddDT from scratch.This work presents a unified pipeline for constructing data-driven Foundation Models for three exemplifying cases: solar-photovoltaic fleets, direct-ink-write additive manufacturing, and laser-powder-bed-fusion additive manufacturing.Although these three systems are conceptually very different, the presented Foundation Model utilizes the flexibility of spatiotemporal graph neural networks (st-GNNs) to apply the same methodology to each case, allowing scientists to focus on their scientific objectives rather than troubleshooting an overwhelmingly detailed physics-based modeling pipeline.