Follow one particle through a beating heart—and let AI infer the river around it

Imagine releasing a tiny dye particle into a flowing stream. A camera might follow the particle as it curls behind a rock. That gives one kind of information: the path of a moving object. A different instrument might describe the speed and direction of water at fixed points across the river. Fluid mechanics has names for these perspectives. The Lagrangian view follows a particle; the Eulerian view observes a field at positions in space. Both describe the same physics, yet combining them accurately is difficult when measurements are scarce.[1]

Researchers publishing in Nature Machine Intelligence on 20 October 2026 introduced TrajectoryFlowNet, a physics-informed neural framework trained to predict both long particle trajectories and the surrounding flow field. Its importance is not that a neural network draws a smooth-looking current. It is that the network is designed to respect physical relationships while integrating measurements that capture only fragments of a moving fluid.[1][2]

An empty map is not an empty river

A laboratory can track a limited number of fluorescent particles, but no instrument can observe every parcel of water or blood simultaneously. Simple interpolation tends to fail where motion twists near boundaries, vortices form or the particle paths cover only a small fraction of the space. Conventional numerical solutions of fluid equations are powerful but can be expensive and sensitive to boundary conditions that are not known exactly. In many practical experiments, the constraints are not computational ambition but missing measurements.[1]

TrajectoryFlowNet mixes observations of movement with relationships drawn from the governing fluid equations. The model learns a continuous representation of the flow while using the tracked paths to constrain that representation. A long predicted particle track should be consistent with the velocities inferred along its route. The continuous field, in turn, should obey the relevant physical conditions. The result is a coupled problem rather than two unrelated prediction models.[1]

Why the blood-flow experiments matter

The paper reports tests on simulated systems including lid-driven cavity flow and flows around complex cylinders. It also includes experimental aortic and ventricular blood-flow cases. Those experiments matter because blood circulation is not an ideal straight pipe: it can accelerate, recirculate, interact with moving boundaries and become difficult to observe. Researchers need better flow reconstruction in order to understand patterns of transport and stress that a few tracked points may miss.[1]

However, experimental validation in a laboratory flow setting is not equivalent to a clinical trial in living patients. The paper does not establish that using TrajectoryFlowNet helps doctors diagnose aortic disease, predict a stroke or change treatment outcomes. The significance lies in reconstructing a richer physical picture from sparse trajectory data under controlled conditions. Accuracy depends on the measurement quality, assumed physics and how closely the test cases represent the situation where the tool is deployed.[1]

A transferable idea beyond cardiovascular imaging

The same dual perspective is useful in other settings. In the atmosphere, a weather balloon moves through a field of air currents. In a factory, particles travel through a mixture as a vessel stirs. In environmental monitoring, drifting markers can reveal where a pollutant might spread. In every case, a limited set of trajectories offers clues about a larger and partly invisible field. A method that can combine the paths and physics may make those clues more useful than either source of information alone.[1]

That possibility should be tested rather than assumed. Turbulent extremes, unknown forces, out-of-distribution boundaries and noisy tracking can expose weaknesses a benchmark misses. Yet there is a compelling principle here: AI need not replace physical theory to be valuable. It can instead become the missing interpreter between measurements made on moving objects and laws that describe motion everywhere around them. The fascinating prediction is not a single particle's destination, but the river that becomes visible because the particle travelled through it.[1]

Report an error or suggest a correction