Industrial Application of a Multi-Disciplinary Design Optimization with Uncertainties to a Pair of Telecommunication Satellites

Anne Gazaix; Fran\c{c}ois Gallard; Gaspard Berthelin (ADS); Lo\"ic Cousin; Matthias De Lozzo; Nicolas Roussouly; Nicolas Sarda (ADS); Olivier Sapin; Xavier Fosse (ADS)

REVIEW 3 major objections 5 minor 10 references

Accounting for uncertainties in the design of two stacked telecom satellites recovers most of the lifetime that conventional safety margins give up, cutting the performance loss by roughly two-thirds while keeping constraint-failure probabi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:35 UTC pith:5F5G627A

load-bearing objection Credible industrial RBMDO case, but the headline 66% figure is provisional because ε is unstated and failure probabilities rest on unverified PCE tails. the 3 major comments →

arxiv 2607.19045 v1 pith:5F5G627A submitted 2026-07-21 math.OC stat.AP

Industrial Application of a Multi-Disciplinary Design Optimization with Uncertainties to a Pair of Telecommunication Satellites

Olivier Sapin , Lo\"ic Cousin , Nicolas Roussouly , Anne Gazaix , Fran\c{c}ois Gallard , Matthias De Lozzo , Xavier Fosse (ADS) , Nicolas Sarda (ADS)

show 1 more author

Gaspard Berthelin (ADS)

This is my paper

classification math.OC stat.AP MSC 90C1590B2565C05

keywords multidisciplinary design optimizationuncertainty quantificationreliability-based design optimizationpolynomial chaos expansiontelecommunication satellitessafety marginsorbital lifetimesensitivity analysis

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that replacing the standard industrial practice of adding fixed safety margins to satellite design constraints with an explicit reliability-based multidisciplinary design optimization (RBMDO) cuts the lifetime penalty of robustness by about two-thirds while keeping constraint-failure probabilities near zero. In the numbers reported, the deterministic optimum gives a 20.43-year orbital lifetime; applying a three-standard-deviation margin to the active constraints yields 19.41 years, whereas the RBMDO solution reaches 20.09 years at essentially the same, near-zero failure probability. The mechanism is to approximate the objective and constraints by polynomial chaos surrogates built from 40 training points in an 18-dimensional uncertain space and then estimate failure probabilities by Monte Carlo sampling on those surrogates at each optimization step. If the result is right, uncertainty-aware design is a practical alternative to margins in industrial satellite sizing, worth roughly four extra months of mission life on a pair of stacked telecom satellites.

Core claim

The authors find that uncertainty can be moved from a post-optimization safety check into the optimization itself without making the computation unaffordable, and that this changes the design outcome. Formulating the propulsion-system MDO as a reliability problem—maximize lifetime subject to a small probability of violating each constraint—yields a feasible design with a 20.09-year lifetime, while the conventional three-sigma-margin design achieves only 19.41 years at comparable reliability; the deterministic, uncertainty-ignored optimum is 20.43 years. The margin route therefore loses 1.02 years of lifetime, the reliability-based route loses 0.34 years, a reduction of about 66%. The active

What carries the argument

The load-bearing machinery is polynomial chaos expansion (PCE): each constraint and the objective are approximated by a truncated expansion on an orthonormal polynomial basis of the 18 Gaussian uncertain inputs, fit by sparse regression on 40 training points from a space-filling design. That surrogate is cheap enough to be sampled 10,000 times by Monte Carlo inside every optimization iteration, giving means, variances, and failure probabilities without calling the full disciplinary simulator. The optimization uses the multidisciplinary-feasible architecture, in which the strongly coupled per-satellite disciplines are solved by fixed-point iteration before constraints are evaluated, with a de

Load-bearing premise

The load-bearing premise is that polynomial surrogates built from 40 training points in an 18-dimensional space represent the constraint tails where failure probabilities near one in a thousand live; if those tails are distorted, both sides of the margin-versus-RBMDO comparison are uncertain.

What would settle it

Re-evaluate the deterministic optimum, the 3-sigma margin design, and the RBMDO optimum using the original high-fidelity disciplinary chain with a Monte Carlo or importance-sampling campaign large enough to resolve failure probabilities around 10^-3 (hundreds to thousands of simulator runs, rather than 40 surrogate training points). If the true failure probability of the RBMDO design is not lower than that of the 3-sigma design, or if it exceeds the requirement epsilon, the paper's central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

At essentially equal failure probability, the RBMDO design keeps 20.09 years of lifetime versus 19.41 years under a 3-sigma margin; the margin approach throws away 1.02 years while RBMDO throws away 0.34 years relative to the 20.43-year deterministic optimum.
Smaller pre-defined margins do not provide a reliability guarantee: 1-sigma margins leave 16-18% failure probabilities on the active constraints and 2-sigma margins leave 1.5-1.9%, while the RBMDO design is at 0.26% or below.
The reliability-based route costs roughly 100 times more per optimization iteration than deterministic MDO, because each iteration builds surrogates from 40 simulator evaluations; that is the price of the high-probability feasibility claim.
Because the two satellites are not coupled to each other, the formulation extends to an arbitrary number of stacked satellites and to multi-objective optimization of all their lifetimes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Editorial extension: the 66% figure is computed from surrogate-derived failure probabilities, and the paper reports no verification against the original high-fidelity simulator; a direct Monte Carlo check at the reported optima would confirm whether the tail probabilities—and hence the margin-versus-RBMDO comparison—survive contact with the full model.
Editorial extension: the acceptable failure probability epsilon is never numerically specified, so the phrase 'high probability' is not fully quantified; the reported 0.26% failure on one constraint at the RBMDO optimum could already violate a stricter requirement.
Editorial extension: the margin baseline tested is an arbitrary k-sigma margin, not a margin calibrated to the same reliability target; a calibrated-margin comparison might narrow the gap, although it would also give up the simplicity that makes the margin practice attractive.
Editorial extension: for systems with heavier tails or stronger interactions among uncertainties, the 40-point, degree-2 surrogate budget would likely need to grow; a tail-focused validation diagnostic would be needed before transferring the cost-benefit numbers to other design problems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Credible industrial RBMDO case, but the headline 66% figure is provisional because ε is unstated and failure probabilities rest on unverified PCE tails.

read the letter

Let me give you my read on arXiv:2607.19045.

The thing to know: this is a genuine industrial RBMDO case study on a stacked pair of GEO telecom satellites, with 18 uncertain variables, strong couplings per satellite, and an internally coherent comparison between deterministic MDO, margin-based designs, and RBMDO. The headline number—66% reduction in performance loss relative to predefined 3σ margins—is directionally plausible but not yet established. The paper doesn't state the reliability threshold ε in Eq. (3), and every failure probability comes from degree-2 PCE surrogates trained on 40 points in 18 dimensions, with no check against the original simulator.

What's actually new and good: no new method, but the quantified benchmark is useful. Table 5 tells a clean story: deterministic optimum at 20.43 yr has ~50% failure on the two active constraints; margins reduce failure as k grows; RBMDO lands at 20.09 yr with 0.26% failure on the electric-propellant constraint and 0.00 on the others; 3σ margins need 19.41 yr for comparable failure. The 100× computational cost is acknowledged. The Sobol' analysis is honest—they identify six low-impact variables and then explain that dropping them didn't save enough to matter, so they kept all 18. The citation pattern is sensible. No fitted constant is folded into the headline, so I don't see circularity.

Soft spots, in proportion. First, ε is never quantified. The text even says ε is "so small" that the constraint plots don't show the margin. Table 5 reports P[elec_propellant_0 ≥ max] = 0.26% for RBMDO and 0.01% for the 3σ design. If the actual required ε is 10⁻³, then the RBMDO solution is infeasible and the comparison collapses. That's a load-bearing omission. Second, tail accuracy. R² > 0.99 is a bulk metric. The claims live at failure probabilities around 10⁻³, and no verification against the Scilab disciplines is reported. Since both comparative panels share the same surrogates, a tail bias would shift both rows in an unknown direction. The 66% figure should be labeled preliminary until at least one true-model tail check is done at the final designs. Third, the 0.00 entries are just <10⁻⁴ with 10⁴ MC draws—fine, but they shouldn't be read as exact zeros. Fourth, units and distribution parameters are withheld for confidentiality, which kills reproducibility. That's a common industry constraint, but it limits the paper's evidential weight.

Who should read it: engineers and tool developers who want a realistic, documented deployment template; not people looking for new methodology. I'd be glad to referee it—or have a student do a careful pass on the PCE validation—after the authors state ε and add a tail verification. I'd also cite it as an industrial data point, with a caveat.

Referee Report

3 major / 5 minor

Summary. The paper reports an industrial application of multidisciplinary design optimization under uncertainty (RBMDO) to a pair of stacked geostationary telecommunication satellites. The deterministic MDO problem maximizes the orbital lifetime of the first satellite subject to mass, electric-orbit-raising duration, propellant, and dry-mass constraints. Uncertainties are modeled as 18 normal random variables, propagated through degree-2 sparse polynomial chaos expansions (PCEs) built from 40 training points, with statistics estimated by 10,000 Monte Carlo draws per evaluation. The paper compares the deterministic optimum, deterministic MDO with kσ margins applied to the active constraints, and a full RBMDO formulation. The central reported result is a 66% reduction in performance loss: compared with the deterministic optimum (20.43 yr), the 3σ-margin design gives 19.41 yr and the RBMDO design gives 20.09 yr. The paper concludes that explicitly accounting for uncertainties is preferable to applying predefined safety margins, while also honestly reporting a roughly 100× per-iteration computational cost increase for the RBMDO approach.

Significance. If the numerical comparison is valid, this is a valuable industrial demonstration of RBMDO on a coupled aerospace system, using open-source tools (GEMSEO, OpenTURNS) and combining sensitivity analysis, PCE surrogates, and reliability-based optimization. The internal comparison in Table 5 is coherent in direction: active constraints at the deterministic optimum show approximately 50% failure, margin-based designs reduce failure as k increases, and the RBMDO design achieves low failure with a smaller lifetime loss. The honest disclosure of the 100× computational cost is a useful practical data point. However, the headline 66% claim rests on failure probabilities estimated from small PCE surrogates in the 10⁻³ tail regime and on an unstated reliability threshold ε; the comparison is not yet tied to verified reliability at the final designs. The paper's contribution is therefore significant but conditional on closing that verification gap.

major comments (3)

[§IV.D, Eq. (3), Tables 4–5] The headline comparison is not made at a stated reliability level. The threshold ε in Eq. (3) is never assigned a numerical value, and Table 5 shows that the RBMDO design has P[elec_propellant_0 ≥ max] = 0.26%, while the 3σ-margin design has 0.01% — a 26-fold difference. The statement in §IV.D that a 3σ margin 'would have been required to achieve the same level of reliability' is therefore not supported by the table. If the intended ε is, for example, 10⁻³, then the RBMDO solution violates its own constraint. Please state ε for each constraint and redo the comparison at equal, explicitly specified reliability targets.
[§III.D–E] All failure probabilities are computed from degree-2 sparse PCEs (up to 30 terms) fitted to 40 training points in an 18-dimensional uncertain space, with 10,000 Monte Carlo samples of the surrogate. Cross-validation R² > 0.99 is reported, but R² measures bulk fit, not tail accuracy, and the claims live near failure probabilities of 10⁻³. No verification against the original Scilab multidisciplinary simulator is reported for either the margin-based or the RBMDO final designs. Since both panels of the comparison use the same surrogate machinery, tail bias could shift both failure probabilities and the resulting lifetimes in an unknown direction; the 66% figure is therefore not yet tied to verified reliability behavior.
[Tables 3–5] Several failure probabilities are reported as exactly 0.00 or 0. With 10,000 Monte Carlo samples, the resolution of a probability estimate is at best about 10⁻⁴, and a value of 0 means only that no violation occurred in that sample. The claim that the RBMDO solution 'guarantees the feasibility of constraints with high probability' is not quantified by these zero entries. Please report the number of violations, confidence intervals, or a sufficiently large MC budget for the final reliability estimates, and describe the precision of the probabilities in Table 5.

minor comments (5)

[Abstract / §IV.C] The margins in Eq. (4) are not 'predefined' in the usual sense: they are chosen as k times the standard deviation estimated at the deterministic optimum. Clarify this in the abstract or in §IV.C to avoid overstating the contrast.
[§III.E heading] Typo: 'Reliablility' should be 'Reliability'.
[Conclusion] Grammar: 'an space system' should be 'a space system'.
[Table 3] The column labeled 'Confidence interval (%)' is not defined. State whether it is a confidence interval for the failure probability, and how it was computed.
[§IV.D / Figure 7] The caption text says 'The same applies to the graph on the left' twice. The second occurrence should refer to the graph on the right.

Circularity Check

0 steps flagged

No significant circularity: the 66% comparison is a numerical outcome of two independent optimization formulations, not a fitted input or self-citation.

full rationale

The central claim—that RBMDO yields a ~66% smaller lifetime loss than a 3σ-margin design—is not forced by construction. The values in Table 5 are outputs of two distinct optimization routes: deterministic MDO with margin-scaled constraint thresholds (Eq. 4) and RBDO with explicit chance constraints (Eq. 3). Neither route folds the other’s output or the headline 66% into its inputs. The unreported reliability threshold ε and the use of degree-2 PCE surrogates for tail probabilities are correctness and falsifiability concerns, not circularity: both compared designs are evaluated with the same surrogate machinery, and the unstated ε does not make Eq. (3) algebraically identical to Eq. (4). References [5] and [6] are normal tooling self-citations (GEMSEO is open source) and are not invoked as a uniqueness theorem or as external validation of the headline. No equation in the paper is definitionally equal to its inputs, and no fitted parameter is relabeled as a prediction. Therefore no specific circular step can be quoted.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities or forces are introduced. The load-bearing 'free' content is: (i) an unstated reliability threshold ε, (ii) undisclosed industrial distribution parameters, and (iii) hand-set surrogate hyperparameters and margin schemes. Everything else is standard UMDO machinery (MDF, PCE, Sobol', RBDO) drawn from the cited literature.

free parameters (4)

Reliability threshold ε for probabilistic constraints = not stated anywhere in the text
Every constraint in Eq. (3) is posed against an unstated ε. The abstract's 'guaranteeing the feasibility ... with high probability' cannot be checked without it, and the RBMDO solution's P[elec_propellant_0] = 3×10⁻³ may exceed a 10⁻³ target.
Means and standard deviations of the 18 normal uncertain variables = confidential Airbus configuration (D_some_data); not disclosed
Nine distributions per satellite are defined in §II.C by 'specified means and standard deviations' whose values come from proprietary configuration files. All failure probabilities inherit these values.
PCE hyperparameters (maximum degree, truncation, training sample size) = degree=2, terms≤30, n=40
Chosen by hand in §III.D 'to ensure accuracy'; every failure estimate in Tables 3–5 is a function of these choices, and they were retained through the RBMDO loop (§III.E).
Margin application scheme in Eq. (4) = k=1,2,3; margins only on eor_duration_0 and elec_propellant_0, with M = k×σ_opt1
The 'generic safety margins' baseline is implemented as pointwise σ multiples at the deterministic optimum on only two constraints; changing this scheme (e.g., margins on all constraints or on the objective) would change the 66% figure.

axioms (4)

domain assumption The 18 uncertain inputs (9 per satellite) are independent Gaussian random variables with known dispersion parameters (§II.C).
Independence and normality are asserted, not validated; correlations among ISP, efficiency, and delta-V dispersions, or between the two satellites' shared launch environment, are excluded.
ad hoc to paper A degree-2 sparse PCE (≤30 terms) trained on 40 points faithfully represents constraint statistics including the 10⁻³ tail regime (§III.D–E).
Quality is checked only via cross-validated bulk R² > 0.99; no comparison of PCE failure probabilities against MC on the true model is reported, and tail accuracy is exactly what the claim needs.
domain assumption The COBYLA (local, derivative-free) solution of the deterministic MDO is a valid reference optimum for defining σ_opt1 and the margin baseline (§III.A, §IV.A).
No multi-start or global verification is reported; a different local optimum would shift both the reference lifetime and the σ estimates the margins are built on.
ad hoc to paper Failure probabilities are estimated to sufficient precision by 10,000 MC samples of the PCEs (Tables 3–5).
For P ≈ 3×10⁻³ the coefficient of variation is ~30%; for reported P = 0.00 the 95% upper bound is ~3×10⁻⁴, so the '0.' entries are censored, not measured guarantees.

pith-pipeline@v1.3.0-alltime-deepseek · 10253 in / 19818 out tokens · 201697 ms · 2026-08-01T13:35:51.525624+00:00 · methodology

0 comments

read the original abstract

In satellite design, it is common practice to add safety margins to the constraints to achieve conservative solutions that are robust to uncertainties. This robustness often comes at the expense of performance and it may be more appropriate to include uncertainties in the definition of the design problem. This work addresses such a challenge by applying techniques of multidisciplinary design optimization under uncertainty to an industrial use case. The latter is a pair of telecommunication satellites launched together in a stacked configuration. Each satellite is a strongly coupled multidisciplinary system while the two satellites are not coupled at all. This use case can be extended to an arbitrary number of satellites. By considering uncertainty quantification techniques such as sensitivity analysis and reliability-based design optimization, this study demonstrates that accounting for uncertainties in the design problem results in a 66% reduction in performance loss compared to the adding of a predefined safety margins, while guaranteeing the feasibility of constraints with high probability.

discussion (0)

Reference graph

Works this paper leans on

10 extracted references

[1]

Multidisciplinary design optimization: a survey of architectures,

Martins, J. R., and Lambe, A. B., “Multidisciplinary design optimization: a survey of architectures,”AIAA journal, Vol. 51, No. 9, 2013, pp. 2049–2075

2013
[2]

Brevault, L., Balesdent, M., Morio, J., et al.,Aerospace system analysis and optimization in uncertainty, Springer, 2020

2020
[3]

Benchmark study of numerical methods for reliability-based design optimization,

Aoues, Y., and Chateauneuf, A., “Benchmark study of numerical methods for reliability-based design optimization,”Structural and multidisciplinary optimization, Vol. 41, No. 2, 2010, pp. 277–294

2010
[4]

Xiu, D.,Numerical methods for stochastic computations: a spectral method approach, Princeton university press, 2010

2010
[5]

GEMS: A Python Library for Automation of Multidisciplinary Design Optimization Process Generation,

Gallard, F., Vanaret, C., Guénot, D., Gachelin, V., Lafage, R., Pauwels, B., Barjhoux, P.-J., and Gazaix, A., “GEMS: A Python Library for Automation of Multidisciplinary Design Optimization Process Generation,”2018 AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, 2018

2018
[6]

Multidisciplinary design optimization under uncertainty: the open source capabilities of GEMSEO,

De Lozzo, M., Laboulfie, C., Sapin, O., Roussouly, N., Gallard, F., Aziz-Alaoui, A., Dechaume, A., and Gazaix, A., “Multidisciplinary design optimization under uncertainty: the open source capabilities of GEMSEO,”6th ECCOMAS Thematic Conference on Uncertainty Quantification in Computational Sciences and Engineering (UNCECOMP), Rhodes, Greece, 2025

2025
[7]

Iooss, B., and Lemaître, P.,A Review on Global Sensitivity Analysis Methods, Springer US, Boston, MA, 2015, pp. 101–122

2015
[8]

Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates,

Sobol, I., “Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates,”Mathematics and Computers in Simulation, Vol. 55, No. 1, 2001, pp. 271–280. The Second IMACS Seminar on Monte Carlo Methods

2001
[9]

Making best use of model evaluations to compute sensitivity indices,

Saltelli, A., “Making best use of model evaluations to compute sensitivity indices,”Computer Physics Communications, Vol. 145, No. 2, 2002, pp. 280–297

2002
[10]

Baudin, M., Dutfoy, A., Iooss, B., and Popelin, A.-L.,OpenTURNS: An Industrial Software for Uncertainty Quantification in Simulation, Springer International Publishing, 2015, p. 1–38. 13

2015

[1] [1]

Multidisciplinary design optimization: a survey of architectures,

Martins, J. R., and Lambe, A. B., “Multidisciplinary design optimization: a survey of architectures,”AIAA journal, Vol. 51, No. 9, 2013, pp. 2049–2075

2013

[2] [2]

Brevault, L., Balesdent, M., Morio, J., et al.,Aerospace system analysis and optimization in uncertainty, Springer, 2020

2020

[3] [3]

Benchmark study of numerical methods for reliability-based design optimization,

Aoues, Y., and Chateauneuf, A., “Benchmark study of numerical methods for reliability-based design optimization,”Structural and multidisciplinary optimization, Vol. 41, No. 2, 2010, pp. 277–294

2010

[4] [4]

Xiu, D.,Numerical methods for stochastic computations: a spectral method approach, Princeton university press, 2010

2010

[5] [5]

GEMS: A Python Library for Automation of Multidisciplinary Design Optimization Process Generation,

Gallard, F., Vanaret, C., Guénot, D., Gachelin, V., Lafage, R., Pauwels, B., Barjhoux, P.-J., and Gazaix, A., “GEMS: A Python Library for Automation of Multidisciplinary Design Optimization Process Generation,”2018 AIAA/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference, 2018

2018

[6] [6]

Multidisciplinary design optimization under uncertainty: the open source capabilities of GEMSEO,

De Lozzo, M., Laboulfie, C., Sapin, O., Roussouly, N., Gallard, F., Aziz-Alaoui, A., Dechaume, A., and Gazaix, A., “Multidisciplinary design optimization under uncertainty: the open source capabilities of GEMSEO,”6th ECCOMAS Thematic Conference on Uncertainty Quantification in Computational Sciences and Engineering (UNCECOMP), Rhodes, Greece, 2025

2025

[7] [7]

Iooss, B., and Lemaître, P.,A Review on Global Sensitivity Analysis Methods, Springer US, Boston, MA, 2015, pp. 101–122

2015

[8] [8]

Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates,

Sobol, I., “Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates,”Mathematics and Computers in Simulation, Vol. 55, No. 1, 2001, pp. 271–280. The Second IMACS Seminar on Monte Carlo Methods

2001

[9] [9]

Making best use of model evaluations to compute sensitivity indices,

Saltelli, A., “Making best use of model evaluations to compute sensitivity indices,”Computer Physics Communications, Vol. 145, No. 2, 2002, pp. 280–297

2002

[10] [10]

Baudin, M., Dutfoy, A., Iooss, B., and Popelin, A.-L.,OpenTURNS: An Industrial Software for Uncertainty Quantification in Simulation, Springer International Publishing, 2015, p. 1–38. 13

2015