Real-time digital twins: unifying bias-aware data assimilation and machine learning

Accurate forecasting and effective control of high-dimensional, nonlinear systems remain key challenges across science and engineering. This seminar explores a new paradigm in real-time digital twins. The digital twin framework unifies data assimilation with machine learning, enabling robust, real-time solutions for modelling, prediction, and control under (i) aleatoric and epistemic uncertainties and (ii) partial observability. This talk covers three key ingredients for real-time twinning: model-error inference, partial observations, and physical control strategies. First, we propose a bias-aware data assimilation framework that combines low-order models with real data while accounting for modelling errors. Second, we demonstrate how data assimilation can be leveraged to integrate partial and noisy observations into machine-learned reduced-order models for time-accurate and numerically stable data-driven models. Third, we explore the integration of partial observations and reduced-order models in reinforcement learning via data assimilation for adaptive decision-making. These approaches overcome limitations of traditional methods—such as the need for full-state measurements and extensive a priori training—while ensuring stability and scalability even in chaotic and turbulent regimes. The presented strategies open new opportunities for advancing digital twins, autonomous systems, and real-time feedback control in complex environments.