New Two-Stage Optimization Puts Rocket Reality Into Satellite Constellation Design

Satellite constellations have become the backbone of modern Earth observation, communications, and responsive monitoring services, but a quiet mismatch has long haunted their design. Engineers typically optimize the orbital pattern of a constellation for coverage first, and only afterward ask whether the chosen satellites can actually be delivered to their assigned orbital planes by real launch vehicles. A new study from the Department of Aerospace Engineering at the Korea Advanced Institute of Science and Technology (KAIST), published in the International Journal of Aeronautical and Space Sciences, closes that gap with a two-stage optimization framework that embeds the physics of deployment directly into the constellation design process itself.

The research, led by Beomjin Gwon and Jaemyung Ahn, tackles a subtle but consequential problem in how multi-satellite launches work. When a single rocket carries many satellites, it injects them into a shared parking orbit rather than directly into their final operational positions. From there, the satellites spread into their target orbital planes by exploiting a natural phenomenon: the J2 perturbation, caused by the oblateness of the Earth, makes orbital planes precess, or slowly rotate, at rates that depend on altitude. Two orbits at different altitudes precess at slightly different speeds, so a satellite that lingers in a parking orbit gradually drifts away from its injection plane. By waiting the right amount of time before firing its thrusters, a satellite can arrive at a target plane without expensive propulsive maneuvers.

The catch is that this drift is slow and the waiting time is finite. The achievable separation in right ascension of the ascending node, or RAAN, the angle that defines where an orbital plane sits around the Earth, is limited by the difference in precession rates between the parking and target orbits multiplied by the available dwell time. Different launchers inject at different epochs and carry different dwell time allowances, so the set of orbital planes each rocket can actually reach is restricted. As the authors point out, not every coverage-attractive constellation pattern is physically reachable within a given deployment window, which means deployment constraints can reshape the feasible design space before the final pattern is ever selected.

To capture this effect mathematically, the framework begins with a repeating ground track seed orbit, an orbital configuration in which the satellite retraces the same path over the Earth’s surface after a fixed cycle. From this seed, the researchers generate a discretized common ground track constellation: a finite family of candidate satellites that share the same ground track and differ only in their coupled RAAN and mean anomaly phasing. This structure yields a powerful computational shortcut. Because all candidates are phase-shifted copies of one another along the same repeat cycle, the temporal access profile of every candidate, a binary record of when it can see a ground target within field-of-view and minimum elevation constraints, can be obtained simply by circularly shifting the seed satellite’s profile.

The deployment physics then enters through what the authors call admissible satellite sets. For each launcher and each discretized injection RAAN option, the reachable RAAN range is approximated as the difference between the parking and target orbit precession rates multiplied by that launcher’s maximum dwell time. Only candidate satellites whose target RAANs fall within this range are deemed deployable by that launcher-injection combination. Aggregating the access contributions of all reachable satellites produces a launcher-level admissible access profile, effectively transforming RAAN reachability into a design-space filter that excludes infeasible patterns during optimization rather than after the fact. The team validated this first-order J2 approximation against full numerical propagation using the EGM2008 gravity model, finding relative errors below 0.66 percent across the tested dwell times, ample accuracy for early design screening.

The optimization itself proceeds in two stages. Stage 1 is an integer linear program that selects the injection RAAN for each launcher, maximizing the worst-case temporal access performance across the simulation horizon while using dwell time margin as a secondary tie-breaking criterion. This screening step identifies injection conditions with favorable coverage potential and deployment slack before any satellite is committed. Stage 2 is a mixed-integer linear program that selects the actual satellites from the admissible pool, assigns them to launchers, and enforces coverage requirements, launcher capacity limits, and the finite deployment window, all while minimizing two conflicting objectives: total deployment cost, comprising satellite and launch costs, and deployment completion time.

Because cost and completion time conflict, the researchers enumerate a Pareto front using an epsilon-constraint strategy with weighted-sum tie breaking. The procedure first finds the minimum-cost and minimum-time extreme points, then iteratively tightens the completion time bound, forcing the solver to move to the next distinct step on the trade-off curve rather than repeatedly returning the same solution. Both stages belong to the class of mixed-integer problems with NP-hard worst-case complexity, and the case studies were run in MATLAB with the Gurobi solver on a high-performance workstation, with each subproblem capped at three hours of wall clock time.

Two case studies demonstrate the framework’s reach. The first designs a Sun-synchronous constellation at roughly 1,257 kilometers altitude to observe Jeju Island, using ten candidate launchers with dwell times ranging from 125 to 356 days and up to eleven satellites per launch. The resulting Pareto front clusters into four-launch and five-launch solutions, with completion times spanning 196.08 to 363.46 days and costs from about 735 million to 843 million dollars under the KSLV-II launcher. Switching to the cheaper Vega C vehicle leaves completion times unchanged but shifts the entire trade-off curve downward in cost, since the orbital mechanics and dwell time constraints are identical. The second case, a lower-inclination LEO constellation at about 826 kilometers targeting New York City with a narrower sensor geometry, requires ten to eleven launches and 76 to 83 satellites, achieving completion times between 86.63 and 132.44 days at costs from roughly 2.32 to 2.54 billion dollars under KSLV-II, or 1.95 to 2.13 billion under Vega C.

The ablation and comparison studies reveal why the integrated approach matters. When the dwell time margin term was stripped from Stage 1, every resulting trade-off solution was dominated by at least one solution from the full framework, and the fast-deployment region of the Pareto front disappeared entirely, with the fastest ablated solution taking 104.38 days versus 86.63 days for the proposed method. Against a conventional coverage-first baseline, which fixed a 36-satellite constellation and then optimized its deployment, the new framework did not uniformly win, but it exposed a far richer set of alternatives: the baseline was marginally faster but 132.70 million dollars more expensive than the fastest proposed solution, while the lowest-cost proposed solution cut total cost by 25.7 percent at the price of 111.83 additional days of deployment time.

The authors position the framework as particularly suited to small and medium near-Earth constellations deployed through multi-satellite launches that exploit J2-driven differential nodal precession, a category that covers much of the modern Earth observation and regional monitoring market. Future work will incorporate orbit-dependent launcher performance, rideshare deployment, explicit transfer maneuver modeling potentially coupled with orbital transfer vehicles, and heterogeneous launch vehicle fleets. For an industry racing to field constellations under tight budgets and tighter schedules, the message is clear: the rocket is not just the delivery service for a constellation, it is a co-designer of it, and treating deployment physics as a first-class citizen of the design space can expose cheaper, faster, and more realistic architectures than coverage optimization alone would ever reveal.

Subject of Research: Deployment-aware optimization of satellite constellation pattern design using J2-driven RAAN reachability and two-stage mixed-integer programming

Article Title: Two-Stage Framework for Deployment-Aware Optimization of Satellite Constellation Pattern Design

Article References: Gwon, B., & Ahn, J. (2026). Two-Stage Framework for Deployment-Aware Optimization of Satellite Constellation Pattern Design. International Journal of Aeronautical and Space Sciences. https://doi.org/10.1007/s42405-026-01298-1

Image Credits: AI Generated

DOI: 10.1007/s42405-026-01298-1

Keywords: satellite constellations, deployment optimization, J2 perturbation, RAAN drift, mixed-integer programming, Pareto front, Sun-synchronous orbit, low Earth orbit, launch vehicles, Earth observation, KAIST, revisit time