A new physics-informed machine learning framework named TrajectoryFlowNet has been introduced to resolve dense, continuous flow fields from sparse tracking data. Published in Nature Machine Intelligence, the study by Wan et al. presents a hybrid approach that bridges the Lagrangian and Eulerian perspectives of fluid dynamics. This method transforms sparse observations into physically accurate, high-resolution maps of complex and chaotic fluid motions. The framework addresses a significant challenge in fluid mechanics, where capturing and mathematically describing complex motions has remained difficult despite centuries of study.
Fluid flows exist ubiquitously in the natural world, ranging from the swirling of atmospheric storm systems to the life-sustaining pumping of blood through the human heart. In examining these flows, scientists often choose between two distinct perspectives: tracking a few isolated, moving particles, known as the Lagrangian perspective, or observing the overall flow passing through a grid, known as the Eulerian perspective. The Lagrangian perspective is similar to sitting in a boat and drifting down a river, where one tracks the specific path or trajectory of individual fluid parcels over time. In contrast, the Eulerian perspective resembles standing on a bridge and watching the river flow past, measuring the velocity and pressure of the fluid at fixed points in space and time.
The framework proposed by Wan et al. is composed of a two-stage deep learning architecture designed to handle the translation between these perspectives. The process begins with sparse data, explicitly taking the initial positions of tracked particles in a fluid flow. Rather than immediately reconstructing the entire flow field, the model first uses a specialized neural network module called the ‘trajectory block’. This module learns the underlying physics of motion to map out the predicted Lagrangian trajectories of the particles from their initial points over time. Once these continuous, physical paths are established, the model feeds the trajectory data into a second neural network referred to as the ‘flow field block’. This secondary network performs a super-resolution step, translating the dense mesh of predicted trajectories into reconstructed Eulerian frames, or flow fields.
By dividing the problem into these two distinct phases, the architecture explicitly mimics the natural physical progression of fluids. This approach avoids treating the mapping between sparse data and high-fidelity flow fields as a naive image-enhancement problem. Previous experimental and numerical approaches have been designed to capture fluid flow behavior from both perspectives in a spatiotemporally high-resolution manner. However, experimental techniques, while capturing ground-truth physical observations, are often limited by sensor resolution and noise, yielding sparse data that are challenging to interpolate. Numerical simulations provide high-resolution spatial flow fields but are often prohibitively expensive computationally and rely on simplifying modeling assumptions that limit their real-world applicability.
The translation of sparse, moving Lagrangian particles or sensors into a high-resolution Eulerian flow field is frequently framed as a sparse sensor reconstruction problem or a super-resolution task in fluid mechanics. Although such super-resolution tasks have been examined for a range of unsteady flow scenarios, current approaches are often limited to recovering instantaneous fields from predefined moving sensors. Temporal prediction remains challenging because of the multiscale nature and chaotic motion inherent to fluid dynamics. Overcoming this limitation to achieve simultaneous instantaneous reconstruction and temporal prediction across both Lagrangian and Eulerian perspectives is an important challenge. Establishing the connection between these dual space–time correlations can characterize how small-scale eddies decorrelate, thereby advancing the understanding of complex flow physics.
The work by Wan et al. provides a direction in sparse sensor reconstruction from the unified Lagrangian–Eulerian perspective. Importantly, the proposed model is not a mere data-driven black box. The model is ‘physics-informed’, meaning that the governing equations of fluid motion, the Navier–Stokes equations, are explicitly provided in the cost function. In addition to assessing the error between predicted and reference trajectories and flow fields, the networks are penalized during training to ensure that the predicted trajectories or flow fields follow the laws of conservation of mass and momentum.
The authors demonstrate this hybrid lens-based technique for a range of flow examples, including left ventricle blood flow. Reconstructing the complex blood flow inside the heart from sparse medical imaging is a critical step forward for personalized cardiovascular diagnostics.
Nonetheless, critical challenges remain on the horizon for the widespread application of this technology. The study covers moderately complex flows, but the model has not yet been applied to highly turbulent, three-dimensional systems, such as the chaotic, multiscale turbulent aerodynamic wake behind an aircraft. The particles in the current model are also assumed to be non-inertial, which can be addressed through the Maxey–Riley formulation, as Zhou and Grauer have successfully demonstrated.
Furthermore, analogous to most data-driven approaches, the proposed model requires retraining for each distinct flow scenario. Though the proposed model demonstrates robustness within each specifically trained case, this reliance on case-by-case retraining highlights the need for the transition from specialized algorithms for a specific case to generalizable architectures. One future direction may be the development of ‘foundational models’ for fluid dynamics, or across diverse physical problems. By pre-training architectures on vast, diverse datasets covering various fluid behaviors, it may be possible to develop universally applicable models that rapidly adapt to unseen situations with zero or minimal fine-tuning.
Equipped with rigorous theoretical frameworks, cutting-edge experimental diagnostics, and high-performance computing paradigms, smart, physics-aware algorithms can support next-generation technologies across a range of science and engineering disciplines.
Subject of Research: Physics-informed machine learning framework for reconstructing fluid dynamics from sparse Lagrangian data
Article Title: Physics-informed machine learning: a hybrid lens for fluid dynamics
Article References: Fukami, K., & Maulik, R. (2026). Physics-informed machine learning: a hybrid lens for fluid dynamics. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01319-4
Image Credits: AI Generated
DOI: 10.1038/s42256-026-01319-4
Keywords: TrajectoryFlowNet, Lagrangian perspective, Eulerian perspective, sparse sensor reconstruction, Navier-Stokes equations, fluid dynamics, deep learning architecture, super-resolution task, left ventricle blood flow, physics-informed neural networks, chaotic fluid motions, data assimilation, Maxey-Riley formulation, aerodynamic wake, personalized diagnostics
News Source: Blake Davidson. (November 21, 2026). Neural Network Maps Fluid Flows from Sparse Particle Data. Scienmag.



