Adaptive Digital Twins: Continuous Subspace Learning for Dynamic Domains

Digital twins can produce unreliable predictions during deployment when their operating environment is uncertain and dynamic. Models are commonly validated for specific environments and conditions, so their training data may not fully represent the test data encountered in real-world scenarios, especially when their operating envelope is unknown or shifting over time. In these situations, extrapolation on out-of-distribution data significantly erodes model accuracy. An example of such a scenario is in predicting atmospheric flight dynamics of a reentry spacecraft, in which the vehicle is subject to unpredictable atmospheric conditions that introduce high uncertainty in flight behavior. To address this, we frame the problem within unsupervised domain adaptation and propose a method to maintain digital twin accuracy in uncertain, evolving environments. We treat the subspace of the test data encountered during deployment as a non-stationary, continuously evolving domain, and adapt a model from training to test data by finding a time-evolving latent space representation of the data. We then evaluate the effectiveness of our method on a synthetic example of a spacecraft digital twin during an aerobraking campaign, where the digital twin must accurately predict thermal loading even when encountering out-of-distribution, extrapolatory atmospheric data.