FlowATC: Flow Matching Model for Aircraft Trajectory Prediction

A new generative route into air traffic prediction

FlowATC is a new aircraft trajectory prediction model built around conditional flow matching, a generative technique that learns how to transport random noise into plausible future trajectories conditioned on an observed flight history [1]. The paper, submitted to arXiv on September 15, 2026, is authored by Mathurin Petit, Emir Torun, Louis Brusset, Jordan Kam and Alexandre M. Bayen, and is listed in the machine learning category cs.LG [1]. A current index page records the same submission and a September 16, 2026 update, which places the work squarely inside the latest 72-hour research window for this subject [3].

The story matters because aircraft trajectory prediction is no longer only a question of drawing a best line through radar or ADS-B points. The paper frames the problem as inherently multimodal: the same observed prefix can lead to more than one operationally plausible future because runway assignment, sequencing decisions and controller instructions may not be visible in position data alone [2]. For a future air traffic management system, that distinction is critical. A deterministic forecast can be accurate on average and still fail to express the range of outcomes that matters for conflict-risk estimation.

FlowATC therefore aims to model a distribution over futures rather than a single future. The authors describe the system as trained exclusively on historical aircraft trajectories, with no route labels and no chart supervision [1]. In other words, the model is not told which departure, arrival or procedure a flight is following. It must infer airspace structure from trajectory data itself.

What FlowATC is trained on

The dataset described in the paper comes from Automatic Dependent Surveillance-Broadcast, or ADS-B, a cooperative surveillance system in which aircraft broadcast state information such as position, altitude, velocity and vertical rate [2]. The authors collected ADS-B data from a live feed over a 60-nautical-mile circular geofence centered on the San Francisco Bay Area [2]. The collection ran from April 10 to April 22, 2026, producing 21,515,794 raw state vectors before segmentation and windowing [2].

After preprocessing, the study reports 1,349,388 trajectory windows of 86 points each, divided into 1,149,245 training windows, 137,127 validation windows and 63,016 test windows [2]. Each window is split into 43 observed points and 43 future points, with the future half spanning about 128 seconds on average [2]. This roughly two-minute prediction horizon is significant: it is short enough to be operationally relevant and long enough for turns, approach decisions and branch-point choices to become visible.

The model does not resample the ADS-B feed to a fixed one-second grid. Instead, it keeps the native irregular sampling interval and gives the network a learned time-delta embedding [2]. That design choice reflects a practical constraint: surveillance feeds are not always perfectly regular, and a usable prediction model should not depend on an idealized data cadence.

The state representation is also deliberately spatial. The paper projects latitude and longitude into a local Cartesian frame centered on San Francisco International Airport and represents each point with six features: three position components and three velocity components [2]. Because the coordinates are absolute rather than aircraft-centered only, the model can learn that certain locations are associated with recurring turns, approach paths or departure flows [2].

The model: trajectory forecasting as inpainting

FlowATC casts trajectory prediction as a sequence-inpainting task [1]. The observed past is treated as clean context, while the unknown future is initialized as noise and denoised into a plausible continuation [2]. The paper represents each 86-step trajectory window as a feature-by-time matrix: the first 43 tokens contain observed aircraft states, and the last 43 tokens are noisy future tokens to be completed [2].

The backbone is a Diffusion Transformer-style architecture with adaptive layer normalization conditioning on generation time [2]. A block-causal attention mask prevents observed tokens from attending to noisy future tokens, while future tokens can attend both to the observed prefix and to one another [2]. This is a subtle but important architectural point. The past remains stable and independent of the sampled noise, while the generated future can maintain internal spatial consistency.

The authors train the same general backbone with two generative objectives: Denoising Diffusion Probabilistic Models and Conditional Flow Matching [1]. In the flow-matching version, the network learns a velocity field that moves Gaussian noise toward the target future trajectory along a path from noise to data [2]. At inference time, Euler integration from time 0 to time 1 in 20 steps produces one sample, and repeating the process with independent noise draws produces multiple candidate futures [2].

This multiple-sample design is not just a technical flourish. For a trajectory with several plausible outcomes, the model can generate a fan of alternatives instead of collapsing them into an average path. The paper emphasizes that averaging across a left turn and a straight continuation can produce a trajectory that belongs to neither real mode [2].

Why route-label-free learning matters

One of FlowATC’s most consequential claims is that it learns without route labels or chart supervision [1]. In aviation, published procedures and navigational fixes encode a large amount of structure, but requiring those labels can make a model harder to deploy across regions, data providers and mixed traffic environments. A system that learns directly from tracks could, in principle, adapt to local operating patterns with less manual annotation.

The San Francisco Bay Area is a useful testbed for that claim because terminal airspace contains repeated structure, branching traffic flows and a mix of commercial and general aviation. The paper reports that FlowATC reproduces known airspace structure around San Francisco International Airport, including the shape of the NIITE FOUR departure procedure, even though it is not supervised with route charts [1]. The full paper further states that the model recovers approach turns, holding patterns and descent profiles from absolute Cartesian coordinates [2].

That result should not be read as proof that FlowATC “understands” procedures in a human operational sense. It is better understood as statistical learning of location-specific trajectory regularities. But that is exactly the point: if historical ADS-B data reliably encodes airspace usage, a sufficiently expressive generative model may capture the local geometry without being handed procedural labels.

Performance against baselines

The headline results compare FlowATC with constant velocity, a deterministic LSTM encoder-GRU decoder, a Trajectron++-style conditional variational autoencoder, and a DDPM version of the same Transformer backbone [1]. At matched parameter count, the paper reports that conditional flow matching outperforms DDPM by 11% to 26% in minADE@20, while both generative objectives surpass the CVAE baseline by 31% to 41% [1].

The detailed table for equal-capacity models uses 1.5 million parameters and evaluates 43 observed points to 43 future points on 63,016 test trajectories [2]. In that comparison, Flow Tiny records a minADE@20 of 390.5 meters and a minFDE@20 of 686.1 meters, compared with 459.9 and 828.1 meters for the DDPM model and 662.9 and 1240.9 meters for the CVAE [2]. The same table also reports lower KDE negative log-likelihood for Flow Tiny at future steps 10, 20 and 43, indicating stronger distributional calibration under the paper’s evaluation protocol [2].

The best-of-K framing is essential to interpreting these numbers. At K=1, a generative model’s single draw is not necessarily better than a deterministic baseline because one random sample is not the same thing as the conditional mean [2]. But when the model is allowed to draw 20 futures, the value of covering the plausible distribution becomes visible. That is much closer to the operational question: not “where is the one predicted point?” but “where could this aircraft plausibly be, and with what density?”

Robustness and operational cost

The paper also tests whether FlowATC remains useful when ADS-B data is temporally decimated, reflecting real-world variation in receiver quality and congestion [2]. When retrained at coarser temporal strides, the model’s minADE@20 degrades gracefully: stride 2 costs 4.9%, while stride 8 costs 44.4% but runs five times faster [2]. This is an important practical result because deployment environments may not all deliver the same feed rate.

Latency is measured on a single NVIDIA RTX 4000 Ada using 20 samples in parallel for flow models and 100 DDIM steps for diffusion models [2]. The paper reports 23.76 ms per K=20 prediction for Flow Tiny, compared with 119.68 ms for Diffusion Tiny [2]. At the larger 20.7-million-parameter scale, Flow Large takes 157.07 ms for K=20, while Diffusion Large takes 773.16 ms [2].

Those numbers support one of the paper’s strongest applied arguments: flow matching may deliver a better speed-accuracy compromise for probabilistic trajectory prediction than diffusion under the tested settings. The authors estimate that, at the mean ADS-B update interval of three seconds, one RTX 4000 Ada could keep forecasts current for roughly 150 aircraft with Flow Tiny or 20 aircraft with Flow Large [2].

Branch points and calibrated maneuver distributions

Beyond displacement errors, the paper evaluates whether FlowATC reproduces maneuver distributions at airspace branch points [2]. The authors build a catalogue of recurring turns from ADS-B trajectories, then compare observed heading-change distributions with sampled outcomes from FlowATC [2]. For 28 branch points with at least 40 test passages, FlowATC’s pooled sampled and observed outcomes have a weighted average 1-Wasserstein distance of 3.3 degrees [2].

The paper also reports that FlowATC predicts a straight-through share of 32.2% against an observed 32.9%, with a correlation of 0.99 across branch points [2]. On an individual-passage basis, Flow Large reduces continuous ranked probability score by 47% relative to a history-blind sampler that knows branch-point distributions but not the specific aircraft history [2]. Its central 80% sampled interval contains the observed outcome for 81.2% of passages, while its central 95% interval contains 91.9% [2].

This branch-point analysis is arguably more operationally meaningful than a single aggregate error table. Airspace conflicts and sequencing problems often emerge at precisely the places where flows split, merge or turn. A model that reproduces both the marginal distribution of maneuvers and a close-to-calibrated conditional distribution for each aircraft is better aligned with decision support than a model optimized only for average displacement.

What remains open

FlowATC is still a research paper, not a certified air traffic control tool. Its training and evaluation focus on one geographic region and one 12-day ADS-B collection period [2]. The authors explicitly train on historical trajectory data only, which is the strength of the route-label-free setup but also a limitation: weather, voice clearances, runway configuration and controller intent are not yet part of the input [1]. The architecture is designed so that additional context such as ATC voice, weather fields or charts could later be appended as tokens, but that remains future-facing within the paper [2].

The publication therefore marks a promising step rather than a finished operational system. Its contribution is to show that conditional flow matching can generate calibrated, multimodal aircraft trajectory distributions from raw historical trajectories and can do so with competitive latency [1]. If future work validates the approach across more airports, seasons, traffic regimes and real-time feeds, FlowATC could become part of the modeling vocabulary for next-generation air traffic management.

Sources from the last 72 hours

1. [1][2609.16528] FlowATC: Aircraft Trajectory Prediction via Flow MatchingSep 15, 2026, 2:21 AM UTC
2. [2]FlowATC: Aircraft Trajectory Prediction via Flow MatchingSep 15, 2026, 2:21 AM UTC
3. [3]FlowATC: Aircraft Trajectory Prediction via Flow Matching - arXiv TrollerSep 16, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.