Industry Partners
Collaboration is at the center of the institute's intellectual priorities, with industrial partnerships enabling transformative impacts across a wide range of application areas. From traditional manufacturing to AI-centric tech companies, the methods developed will allow for significant advancements in the AI engineering sciences. Moreover, it will create a virtuous cycle of workforce development in AI for dynamics.
The Aerospace Corporation is a Federally Funded Research Development Center dedicated to supporting the space enterprise and acts as an unbiased technical advisor to the government on complex problems. At a time when space is proliferating and technology is rapidly changing, The Aerospace Corporation is examining the role that machine learning can play in the space domain while still evaluating potential risks of automation in high-consequence environments.
Nick Zolman’s research at The Aerospace Corporation is focused on finding ways to incorporate extensive engineering and physics expertise into machine learning algorithms so that they can operate in the low-data limit. Examples of his work include building lithium-ion battery digital twin models to estimate battery health on-orbit and developing methods to extract low-dimensional representations of dynamics from high-dimensional scenes in the absence of ground truth state-measurements.
Nick Zolman’s research at The Aerospace Corporation is focused on finding ways to incorporate extensive engineering and physics expertise into machine learning algorithms so that they can operate in the low-data limit. Examples of his work include building lithium-ion battery digital twin models to estimate battery health on-orbit and developing methods to extract low-dimensional representations of dynamics from high-dimensional scenes in the absence of ground truth state-measurements.
Aero-Optics and Beam Control
Aero-optical beam control relies on the development of low-latency forecasting techniques to quickly predict wavefronts aberrated by the Turbulent Boundary Layer (TBL) around an airborne optical system, and its study applies to a multi-domain need from astronomy to microscopy for high-fidelity laser propagation. We leverage the forecasting capabilities of the Dynamic Mode Decomposition (DMD) -- an equation-free, data-driven method for identifying coherent flow structures and their associated spatiotemporal dynamics -- in order to estimate future state wavefront phase aberrations to feed into an adaptive optic (AO) control loop. We specifically leverage the optimized DMD (opt-DMD) algorithm on a subset of the Airborne Aero-Optics Laboratory Transonic (AAOL-T) experimental dataset, characterizing aberrated wavefront dynamics for 23 beam propagation directions via the spatiotemporal decomposition underlying DMD. Critically, we show that opt-DMD produces an optimally de-biased eigenvalue spectrum with imaginary eigenvalues, allowing for arbitrarily long forecasting to produce a robust future-state prediction, while exact DMD loses structural information due to modal decay rates. This data-driven decomposition of the aero-optics interaction allows for rapid and robust methods for integrating into machine learning control algorithms more broadly. Continuing work is aimed at characterizing data-fusion and optimal sensing for improving latency for control. This work is led by graduate student Shervin Sahba along with Chris Wilcox, Austin McDaniel and Benjamin Schaeffer (Kirkland Air Force Research Laboratories, Albuquerque NM) and J. Nathan Kutz and Steven Brunton.
Aero-optical beam control relies on the development of low-latency forecasting techniques to quickly predict wavefronts aberrated by the Turbulent Boundary Layer (TBL) around an airborne optical system, and its study applies to a multi-domain need from astronomy to microscopy for high-fidelity laser propagation. We leverage the forecasting capabilities of the Dynamic Mode Decomposition (DMD) -- an equation-free, data-driven method for identifying coherent flow structures and their associated spatiotemporal dynamics -- in order to estimate future state wavefront phase aberrations to feed into an adaptive optic (AO) control loop. We specifically leverage the optimized DMD (opt-DMD) algorithm on a subset of the Airborne Aero-Optics Laboratory Transonic (AAOL-T) experimental dataset, characterizing aberrated wavefront dynamics for 23 beam propagation directions via the spatiotemporal decomposition underlying DMD. Critically, we show that opt-DMD produces an optimally de-biased eigenvalue spectrum with imaginary eigenvalues, allowing for arbitrarily long forecasting to produce a robust future-state prediction, while exact DMD loses structural information due to modal decay rates. This data-driven decomposition of the aero-optics interaction allows for rapid and robust methods for integrating into machine learning control algorithms more broadly. Continuing work is aimed at characterizing data-fusion and optimal sensing for improving latency for control. This work is led by graduate student Shervin Sahba along with Chris Wilcox, Austin McDaniel and Benjamin Schaeffer (Kirkland Air Force Research Laboratories, Albuquerque NM) and J. Nathan Kutz and Steven Brunton.
Advanced Manufacturing. Manufacturing is a highly complex and dynamic process, involving the coordination and merging of several elaborate and precisely timed stages. In a modern manufacturing environment, tremendous volumes of data are being generated, stored, and analyzed to improve process quality, reliability, and efficiency. For example, a Boeing 787 comprises 2.3 million parts that are sourced from around the globe and assembled in an extremely complex and intricate manufacturing process, resulting in vast multimodal data from supply chain logs, video feeds in the factory, inspection data, and hand-written engineering notes. After assembly, a single flight test will collect data from 200,000 multimodal sensors, including asynchronous signals from digital and analogue sensors, including strain, pressure, temperature, acceleration, and video. Thus, big data is presently a reality in modern aerospace engineering, and the field is ripe for AI/ML.
There are several opportunities to leverage ML to improve manufacturing processes. Several areas of high priority include: part standardization; automation and robotics; streamlined assembly, including reduced measurements, processing, and inspection, towards just-in-time manufacturing; supply chain management; material design and fabrication; and non-destructive inspection. Working with our partners at Boeing, we have recently developed a sparse sensor algorithm to dramatically reduce measurements required for precision aircraft assembly, streamlining the manufacturing process. This technology is currently in production on the 777X and 787.
Thanks in part to a $10 million donation from The Boeing Company, the University of Washington is constructing a new Interdisciplinary Engineering Building, will provide an academic home for all undergraduate engineering students, advancing engineering education at the UW and adding new, state-of-the-art spaces for students to learn and collaborate.
There are several opportunities to leverage ML to improve manufacturing processes. Several areas of high priority include: part standardization; automation and robotics; streamlined assembly, including reduced measurements, processing, and inspection, towards just-in-time manufacturing; supply chain management; material design and fabrication; and non-destructive inspection. Working with our partners at Boeing, we have recently developed a sparse sensor algorithm to dramatically reduce measurements required for precision aircraft assembly, streamlining the manufacturing process. This technology is currently in production on the 777X and 787.
Thanks in part to a $10 million donation from The Boeing Company, the University of Washington is constructing a new Interdisciplinary Engineering Building, will provide an academic home for all undergraduate engineering students, advancing engineering education at the UW and adding new, state-of-the-art spaces for students to learn and collaborate.
Promoting Optimal Sparse Sensing and Sparse Learning for Nuclear Digital Twins
In collaboration with INL, we are optimizing sensor placements for nuclear digital twins given spatial constraints and hostile operating conditions for sensor installation. In this setting, sensors critically enable communication between the virtual/digital twin (ROMs and simulations) of the Transient Water Irradiation System in TREAT (TWIST) prototype, and the experimental/physical facility. The TWIST prototype is a multi-purpose test rig that will simulate transient loss of coolant, out-of-pile, to study thermal-hydraulics behavior of an identical irradiation rig for the INL Transient Reactor Test Facility (TREAT). Our algorithms, equipped with physics simulations and ROMs at the design stage, are extended to adaptively optimize sensors with emerging spatial constraints imposed by the environment. The resulting sensor-based reconstructions of reactor flow fields minimize error, provide probabilistic estimates of noise-induced uncertainty, and ultimately will be used for robust monitoring and control of the TWIST Capsule.
In collaboration with INL, we are optimizing sensor placements for nuclear digital twins given spatial constraints and hostile operating conditions for sensor installation. In this setting, sensors critically enable communication between the virtual/digital twin (ROMs and simulations) of the Transient Water Irradiation System in TREAT (TWIST) prototype, and the experimental/physical facility. The TWIST prototype is a multi-purpose test rig that will simulate transient loss of coolant, out-of-pile, to study thermal-hydraulics behavior of an identical irradiation rig for the INL Transient Reactor Test Facility (TREAT). Our algorithms, equipped with physics simulations and ROMs at the design stage, are extended to adaptively optimize sensors with emerging spatial constraints imposed by the environment. The resulting sensor-based reconstructions of reactor flow fields minimize error, provide probabilistic estimates of noise-induced uncertainty, and ultimately will be used for robust monitoring and control of the TWIST Capsule.
UW and MERL scientists collaborate on reduced-order modeling of physical systems for forecasting, control, and sensing applications. In particular, a framework for incorporating knowledge of physics into neural-network-based reduced-order models was developed by Aleksei Sholokhov (UW) and Yuying Liu (UW) under the guidance of Hassan Mansour (MERL), Joshua Rapp (MERL), and Saleh Nabi (MERL). They showed that resulting models predict more accurately, extrapolate better in unforeseen scenarios, and less sensitive to noise. This work already resulted in several papers and remains an active research direction.
UW Ph.D. Student
Applied Mathematics
Power Grid Line Fault Identification
The effort with PNNL is focused on the power grid fault line identification problem through the lens of dynamical systems and Koopman operators. The Koopman operator theory provides us the foundation to represent the dominant and transient dynamical signatures of nonlinear disturbances in a power grid system in the form of linear Koopman modes. Using optimized DMD and time-delayed, Hankelized embeddings, we can approximate the Koopman modes with a stable and robust algorithm. We approach the fault line identification problem as a classification task and propose a decision tree-based AdaBoost ensemble model in order to promote sparse PMU placement. This approach gives satisfactory classification results using very few PMU measurements with simulations on the IEEE 68-bus power system, as well as a more complex IEEE 145-bus power system. The problem emphasizes the use of sparse PMU measurements for fault line classification, so the interpretability of the model plays an important role in our choice of AdaBoost model. Classifiers with more complicated structures, such as the neural nets, shall be considered carefully in future work for improved performance with model interpretability in mind. The work is led by graduate student Frank Mei (AI Institute) is in collaboration with David Solano-Barajas, Alex Tartakovsky, and J. Nathan Kutz.
The effort with PNNL is focused on the power grid fault line identification problem through the lens of dynamical systems and Koopman operators. The Koopman operator theory provides us the foundation to represent the dominant and transient dynamical signatures of nonlinear disturbances in a power grid system in the form of linear Koopman modes. Using optimized DMD and time-delayed, Hankelized embeddings, we can approximate the Koopman modes with a stable and robust algorithm. We approach the fault line identification problem as a classification task and propose a decision tree-based AdaBoost ensemble model in order to promote sparse PMU placement. This approach gives satisfactory classification results using very few PMU measurements with simulations on the IEEE 68-bus power system, as well as a more complex IEEE 145-bus power system. The problem emphasizes the use of sparse PMU measurements for fault line classification, so the interpretability of the model plays an important role in our choice of AdaBoost model. Classifiers with more complicated structures, such as the neural nets, shall be considered carefully in future work for improved performance with model interpretability in mind. The work is led by graduate student Frank Mei (AI Institute) is in collaboration with David Solano-Barajas, Alex Tartakovsky, and J. Nathan Kutz.
