acceptodds
Under review as a conference paper at ICLR 2027

Learning Probabilistic Airspace Demand from Microscopic Aircraft Interaction Graphs

Abstract

Air traffic flow prediction is essential for proactive air traffic management, yet existing approaches predominantly forecast future regional flows from aggregated time-series histories. Such aggregation can obscure the states and interactions of individual aircraft that directly shape near-term traffic. We investigate whether the current airspace snapshot can instead support probabilistic demand forecasting without historical flow sequences. We introduce AEROGRAPH, which represents the variable aircraft population as a dynamic multi-relational graph with physically typed interactions and regional task nodes, and predicts discrete distributions over 15-minute aircraft counts. Across two real-world airports, microscopic snapshot-based models consistently outperform time-series models. Ablation studies identify kinematic interactions and regional task structure as the most consistent contributors. Under bidirectional cross-airport zero-shot transfer, AEROGRAPH retains its advantage under spatial-temporal domain shift. These results show that airspace snapshot modeling provides a strong alternative to aggregated time-series histories for near-term airspace demand forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Air traffic flow forecasting, probabilistic count forecasting, microscopic aircraft-state representation, dynamic multi-relational graphs, graph neural networks, predictive uncertainty.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YWqAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper constructs a dynamic multi-relational graph in which aircraft are represented as nodes connected through four types of relations: spatial, kinematic, intent, and order. The framework is designed for two region-specific prediction tasks. The experiments are comprehensive, and the proposed relations have clear practical interpretations. However, the methodological novelty is limited, as the framework mainly combines existing graph learning techniques.

Critical strengths
  1. Using aircraft states to predict future regional demand is meaningful compared with traditional time-series prediction, as it allows the model to capture interactions among aircraft.
  2. The four types of relations including spatial, kinematic, intent, and order have clear physical meaning.
  3. The experiments are comprehensive and demonstrate substantial effort.
Critical weaknesses
  1. The methodological innovation is limited. The multi-relation graph, relation-specific message passing, attention gating and task-conditioned pooling are all established techniques, and the proposed framework mainly combines them. For ICLR that clearly expects algorithm or method novelty, authors may consider incorporate physical constraints into the evolving graph relations to better reflect aircraft interactions. I would encourage the authors to consider submitting this work to a domain-specific journal, where its practical contributions may be better appreciated, rather than ICLR, which places greater emphasis on methodological novelty.

  2. The authors primarily use adjacency matrices to construct a supergraph. However, the edges weights can carry important information. Adding physical constraints to model relation dynamics on edges could improve the novelty.

  3. The four types of relations are constructed according to the predefined rules in Table 6. These rules are reasonable domain knowledge, but extensive sensitivity analysis is not particularly necessary or informative. introducing some mechanisms for adaptive relation construction would be more suitable for a submission to ICLR.

Questions for authors

see weaknesses

2 comments
AuthorsOpenReview ↗

Thank you for your time and insightful comments

Thank you for taking the time to review our work. We will give your suggestion of a domain-specific journal serious thought. We are also grateful for your concrete suggestions on incorporating physical constraints into the graph and learning relations adaptively.

We especially appreciate the care you took in considering where this work might find a receptive audience. Kind regards.

Thanks for your reply. I will keep my assessment.

Reviewer_2J7ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper links the physically meaningful relationships between aircraft through four physically motivated relation types: spatial, kinematic, intent, and floworder interactions. At the same time, it designs aircraft nodes to be connected with both an airspace approach region (AP) and control region (AR) task nodes, characterizing the current aircraft swarm as a dynamic multi-relational graph. This paper proposes the AEROGRAPH model, capable of outputting a discrete probability distribution for the number of aircraft in each area over the next 15 minutes. Trained and tested on more than 510,000 samples from two domestic airports, the model yields promising prediction results.

Critical strengths

1.The authors characterize the physical connections between aircraft through four physically motivated relation types, while simultaneously accounting for the relationships between aircraft and both approach region and control region. This design effectively incorporates objective states into the model, thereby improving predictive performance.

2.The dataset is sufficiently large, and bidirectional zero-shot transfer experiments are conducted. This work represents a valuable exploration of future data utilization.

Critical weaknesses
  1. The annotation style of individual samples is not presented, making it impossible to examine the specific annotations of a single sample example and evaluate their reasonableness.

  2. It is suggested to conduct analysis on more similar tasks in order to further highlight the unique features that distinguish this work from other domain prediction tasks.

Questions for authors

1.Would it be possible for readers to request access to the annotation details of individual samples? If permitted, how should such a request be submitted?

2.Have the authors considered the potential impact of different airspace control rules across different countries on the model's prediction accuracy? Specifically, how does the effect of varying airspace control rules compare to that of kinematic relations?

3.This article predicts the number of duplicate occurrences within a unit time. If the aircraft pass one after another or simultaneously, there will be different instantaneous loads. Therefore, this prediction needs additional analysis or experiments compared to the actual space traffic control.

1 comment
AuthorsOpenReview ↗

Thank you for your time and comments

Thank you for your encouraging evaluation and thoughtful suggestions. We appreciate your recognition of the physically motivated graph representation and the experimental evaluation.

We will clarify how sample labels are derived from aircraft tracks, illustrate this with a worked example, and explain whether and how readers can request sample-level details, subject to data-sharing permissions. We will also clarify how this task relates to similar forecasting tasks. The relative effects of local control rules and kinematic relations remain an open question that would benefit from evaluation across airspaces.

We appreciate your point about instantaneous load. The same number of aircraft entering within an interval can lead to different peak loads, depending on when they enter and how long they remain. We will make this distinction clearer and discuss how instantaneous occupancy could complement interval demand in operational evaluation.

Thank you again for taking the time to help us improve the paper. Kind regards.

Reviewer_5GAgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper use a dynamic graph with 4 aircraft-aircraft relations and two aircraft-to-region task edges. The motivation is to use current airspace snapshot to predict the distribution of near-future regional demand without relying on historical demand time series.

Critical strengths

The relations have practical meaning and the experiments are comprehensive.

Critical weaknesses
  1. The motivation is somehow wired. Is there any support for the claim "but aggregation removes much of the aircraft-level information available at the current time"? Usually, the state at t=0 is also used as a historical state in most architectures. If you believe the current state is important, you can use it as an anchor or put high weights on current state. I cannot support the motivation of this paper.
  2. I see little new in architecture or technology. Each block is very common, different kinds of MLP for feature representation, the dynamical graph is very common graph. Edge features are concatenated, without the relation-induced mechanisms. Authors write 5 subsections in section 4, but they are just a traditional MPNN pipeline.
  3. The experiments seem comprehensive, but most baselines are completely out-dated. Graph is a fast evolving area, baselines of 3 years ago are really old. Authors compares a lot of baselines published before 2020.
Questions for authors

see the weakness

3 comments
AuthorsOpenReview ↗

Thank you for your time and comments

Thank you for your evaluation and thoughtful comments.

We could have explained our motivation more clearly. Even at the current time step, the same aircraft count can reflect different aircraft positions and headings. This is why we use aircraft-level observations for near-term demand forecasting. We agree that the architecture uses standard message-passing components.

By now, we pay much more attention on its practical value and might weaken the architecture novelty.

Thank you for taking the time to review our work. Kind regards.

Thanks for this reply. This comment is not to belittle your work. Instead, I can see a lot of effort you have done in this work. Good figures and complete experiments. However, I think authors should consider about the underling dynamics, especially the interaction dynamics, physical dynamics, instead to claim a combination of loss to fit these underlying mechanism. In fact, aircraft interaction has much similarity to traffic interaction. The motion of vehicles, including position and headings, is also rapidly changing. Furthermore, vehicle include more HVs, lower automated control penetration. Thus, the dynamics is more complex than aircraft. I think authors can consider learning from vehicle interactions. Especially, in most recent work, the macroscopic model has been transferred from traffic to aircraft. The fundamental diagram has been utilized to aircraft. You can try from this perspective.

AuthorsOpenReview ↗

Thanks for your advice

Thanks for your reply. We totally agree with your opinion.

In fact, incorporating only the current (i.e., t=0) aircraft-level configuration, aims especially for modeling the most current interaction dynamics, which we also think this would be more effecitive than taking the historical time series as input. While the claim of a combination of loss might make our aim vague.

We will also try more macroscopic model from recent traffic studies to improve this study. Thank you for these advice.

Best,

The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.