an author's https://oatao.univ-toulouse.fr/27577
Caleb, Thomas and Lizy-Destrez, Stéphanie Can Uncertainty Propagation Solve The Mysterious Case of Snoopy ? (2021) In: COMET ORB 2021, 9 February 2021 - 10 February 2021 (Virtual event, France).
Uncertainty propagation with TDA
The mysterious case of Snoopy
Thomas Caleb
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics
Preventing collisions between spacecraft Avoiding rendezvous failures
Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics
Preventing collisions between spacecraft Avoiding rendezvous failures
Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics
Preventing collisions between spacecraft
Avoiding rendezvous failures Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics Preventing collisions between spacecraft
Avoiding rendezvous failures
Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics Preventing collisions between spacecraft
Avoiding rendezvous failures
Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics Preventing collisions between spacecraft
Avoiding rendezvous failures Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Introduction
With every measure comes an uncertainty
Uncertainty propagation is crucial in orbital mechanics Preventing collisions between spacecraft
Avoiding rendezvous failures Forecasting collisions with asteroids
Growing number of spacecraft =⇒ Need to quantify these uncertainties to avoid major collisions (Iridium 33 And Cosmos 2251)
Objective
Implementing a propagator in C++, able to compute trajectories and quantify their uncertainties, especially for long propagations
Table of contents I
1 Introduction
2 TDA and Dynamics
3 Case studies
Introduction to TDA
Taylor Differential Algebra (TDA): Assimilation of functions of n variables with their Taylor expansions of order k [Berz 1999]
Similar as the processing of real numbers into floating points
Figure:Analogy between floats and TDA (from [Armellin 2010])
Libraries implementing TDA:
Audi (C++ and Python) by D. Izzo and F. Biscani [Izzo 2018] DACE (C++) by Politecnico di Milano [Massari 2017]
Introduction to TDA
Taylor Differential Algebra (TDA): Assimilation of functions of n variables with their Taylor expansions of order k [Berz 1999]
Similar as the processing of real numbers into floating points
Figure:Analogy between floats and TDA (from [Armellin 2010])
Libraries implementing TDA:
Audi (C++ and Python) by D. Izzo and F. Biscani [Izzo 2018] DACE (C++) by Politecnico di Milano [Massari 2017]
Introduction to TDA
Taylor Differential Algebra (TDA): Assimilation of functions of n variables with their Taylor expansions of order k [Berz 1999] Similar as the processing of real numbers into floating points
Figure:Analogy between floats and TDA (from [Armellin 2010])
Libraries implementing TDA:
Audi (C++ and Python) by D. Izzo and F. Biscani [Izzo 2018] DACE (C++) by Politecnico di Milano [Massari 2017]
Introduction to TDA
Taylor Differential Algebra (TDA): Assimilation of functions of n variables with their Taylor expansions of order k [Berz 1999] Similar as the processing of real numbers into floating points
Figure:Analogy between floats and TDA (from [Armellin 2010])
Libraries implementing TDA:
Introduction to TDA
Taylor Differential Algebra (TDA): Assimilation of functions of n variables with their Taylor expansions of order k [Berz 1999] Similar as the processing of real numbers into floating points
Figure:Analogy between floats and TDA (from [Armellin 2010])
Libraries implementing TDA:
Audi (C++ and Python) by D. Izzo and F. Biscani [Izzo 2018]
DACE (C++) by Politecnico di Milano [Massari 2017]
Propagating uncertainties using TDA
The following equation needs to be solved numerically : (
˙y (t) = f (y (t), t) y (0) = y0
Initial condition [y0]is a polynomial : [y0](
δ
y0) =y0+δ
y0The whole uncertainty space is represented by [y0]
Integration methods (Euler’s method here) can be adapted to: [yn+1] = [yn] +h · f ([yn],tn) +O(h2) (1)
The sequence of polynomials ([yn])n∈N is an expansion of the flow
Propagating uncertainties using TDA
The following equation needs to be solved numerically : (
˙y (t) = f (y (t), t) y (0) = y0
Initial condition [y0]is a polynomial : [y0](
δ
y0) =y0+δ
y0The whole uncertainty space is represented by [y0]
Integration methods (Euler’s method here) can be adapted to: [yn+1] = [yn] +h · f ([yn],tn) +O(h2) (1)
The sequence of polynomials ([yn])n∈N is an expansion of the flow
around y0
Propagating uncertainties using TDA
The following equation needs to be solved numerically : (
˙y (t) = f (y (t), t) y (0) = y0
Initial condition [y0]is a polynomial : [y0](
δ
y0) =y0+δ
y0The whole uncertainty space is represented by [y0]
Integration methods (Euler’s method here) can be adapted to: [yn+1] = [yn] +h · f ([yn],tn) +O(h2) (1)
The sequence of polynomials ([yn])n∈N is an expansion of the flow
Propagating uncertainties using TDA
The following equation needs to be solved numerically : (
˙y (t) = f (y (t), t) y (0) = y0
Initial condition [y0]is a polynomial : [y0](
δ
y0) =y0+δ
y0The whole uncertainty space is represented by [y0]
Integration methods (Euler’s method here) can be adapted to: [yn+1] = [yn] +h · f ([yn],tn) +O(h2) (1)
The sequence of polynomials ([yn])n∈N is an expansion of the flow
around y0
Propagating uncertainties using TDA
The following equation needs to be solved numerically : (
˙y (t) = f (y (t), t) y (0) = y0
Initial condition [y0]is a polynomial : [y0](
δ
y0) =y0+δ
y0The whole uncertainty space is represented by [y0]
Integration methods (Euler’s method here) can be adapted to: [yn+1] = [yn] +h · f ([yn],tn) +O(h2) (1)
Domain splitting methods I
How to increase the precision on the Taylor approximation ?
Increasing the order k of the TDA =⇒ Dramatic growth of the computational time [Wittig 2015, Armellin 2010]
Quantification of the error made by the approximation, since: f (x +
δ
x ) = f (x ) + f(1)(x )δ
x + ... +f(k )(x )
k !
δ
xk +o
δ
xk +1 (2)With x the expansion point and
δ
x the uncertainty on x Then the error is:(
δ
x ) ≤ C f (k +1)(x ) (k + 1)! |δ
x | k +1 (3) With C > 0Domain splitting methods I
How to increase the precision on the Taylor approximation ?
Increasing the order k of the TDA =⇒ Dramatic growth of the computational time [Wittig 2015, Armellin 2010]
Quantification of the error made by the approximation, since: f (x +
δ
x ) = f (x ) + f(1)(x )δ
x + ... +f(k )(x )
k !
δ
xk +o
δ
xk +1 (2)With x the expansion point and
δ
x the uncertainty on x Then the error is:(
δ
x ) ≤ C f (k +1)(x ) (k + 1)! |δ
x | k +1 (3) With C > 0Domain splitting methods I
How to increase the precision on the Taylor approximation ? Increasing the order k of the TDA =⇒ Dramatic growth of the computational time [Wittig 2015, Armellin 2010]
Quantification of the error made by the approximation, since: f (x +
δ
x ) = f (x ) + f(1)(x )δ
x + ... +f(k )(x )
k !
δ
xk +o
δ
xk +1 (2)With x the expansion point and
δ
x the uncertainty on xThen the error is:
(
δ
x ) ≤ C f (k +1)(x ) (k + 1)! |δ
x | k +1 (3) With C > 0Domain splitting methods I
How to increase the precision on the Taylor approximation ? Increasing the order k of the TDA =⇒ Dramatic growth of the computational time [Wittig 2015, Armellin 2010]
Quantification of the error made by the approximation, since: f (x +
δ
x ) = f (x ) + f(1)(x )δ
x + ... +f(k )(x )
k !
δ
xk +o
δ
xk +1 (2)With x the expansion point and
δ
x the uncertainty on xThen the error is:
(
δ
x ) ≤ C f (k +1)(x ) |δ
x |k +1 (3)Domain splitting methods II
If ∆x > 0 is the size of the uncertainty space, then:
(
δ
x ) ≤ C f (k +1)(x ) (k + 1)! ∆x k +1= max (4)Splitting the domain in half reduces the error
(
δ
x ) ≤ max2k +1 (5)
Domain splitting methods II
If ∆x > 0 is the size of the uncertainty space, then:
(
δ
x ) ≤ C f (k +1)(x ) (k + 1)! ∆x k +1= max (4)Splitting the domain in half reduces the error
(
δ
x ) ≤ maxDomain splitting methods III
Figure:Domain Splitting principle (inspired by [Wittig 2015])
Domain splitting methods III
Figure:Domain Splitting principle (inspired by [Wittig 2015])
Domain splitting methods III
Figure:Domain Splitting principle (inspired by [Wittig 2015])
Domain splitting methods III
Figure:Domain Splitting principle (inspired by [Wittig 2015])
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series
The error can be lowered by increasing the order, or by splitting the initial domain as needed
Why choosing TDA propagation ?
Main properties
Uncertainty propagation
All ODE solvers can be adapted to TDA
The resulting trajectory can be evaluated on all the uncertainty space. Only one propagation is needed
A single TDA-propagation takes more time than a classic propagation
Estimation error
An approximation error is made when expanding to Taylor series
The error can be lowered by increasing the order, or by splitting the initial domain as needed
Ephemeris Model and SRP
The acceleration−→
γ
is written as: − →γ
= X body ∈bodies − →γ
body+−→γ
SRP (6)Position of attracting bodies delivered by SPICE [Folkner 2014]: − →
γ
body =μ
body − → r body −−→r − → r body −−→r 3 (7)Solar Radiation Pressure (SRP) expressed as [Georgevic 1971]: − →
γ
SRP= − CRKSRPS m · − → r Sun−−→r − → r Sun− − → r 3 (8)Ephemeris Model and SRP
The acceleration−→
γ
is written as: − →γ
= X body ∈bodies − →γ
body+−→γ
SRP (6)Position of attracting bodies delivered by SPICE [Folkner 2014]: − →
γ
body =μ
body − → r body −−→r − → r body −−→r 3 (7)Solar Radiation Pressure (SRP) expressed as [Georgevic 1971]: − →
γ
SRP= − CRKSRPS m · − → r Sun−−→r − → r Sun− − → r 3 (8)Ephemeris Model and SRP
The acceleration−→
γ
is written as: − →γ
= X body ∈bodies − →γ
body+−→γ
SRP (6)Position of attracting bodies delivered by SPICE [Folkner 2014]: − →
γ
body =μ
body − → r body −−→r − → r body −−→r 3 (7)Solar Radiation Pressure (SRP) expressed as [Georgevic 1971]: −
Modelling SRP uncertainties
The uncertainties of SRP parameters are gathered in a single variable: [−→
γ
SRP] = −[CR0 ] · KSRPS m · [−→r ] [r ]3 (9) With:δ
CR0 CR0 = sδ
CR CR 2 +δ
S S 2 +δ
m m 2 (10)The expression of the acceleration can be simplified with: [
μ
Sun] =μ
Sun− [CR0 ] ·KSRPS
m (11)
Offers a speed-up of at least 2compared to the modeling of the uncertainties on each variable
Modelling SRP uncertainties
The uncertainties of SRP parameters are gathered in a single variable: [−→
γ
SRP] = −[CR0 ] · KSRPS m · [−→r ] [r ]3 (9) With:δ
CR0 CR0 = sδ
CR CR 2 +δ
S S 2 +δ
m m 2 (10) The expression of the acceleration can be simplified with:[
μ
Sun] =μ
Sun− [CR0 ] ·KSRPS
m (11)
Offers a speed-up of at least 2compared to the modeling of the uncertainties on each variable
Modelling SRP uncertainties
The uncertainties of SRP parameters are gathered in a single variable: [−→
γ
SRP] = −[CR0 ] · KSRPS m · [−→r ] [r ]3 (9) With:δ
CR0 CR0 = sδ
CR CR 2 +δ
S S 2 +δ
m m 2 (10) The expression of the acceleration can be simplified with:[
μ
Sun] =μ
Sun− [CR0 ] ·KSRPS
m (11)
Offers a speed-up of at least 2compared to the modeling of the uncertainties on each variable
Comet C/2013 A1 (Siding Spring)
Propagation of the trajectory of the comet Siding Spring
Similar trajectory as the one of Snoopy [Farnocchia 2016]
Figure:Trajectory of Siding Spring (NASA)
Position available on SPICE for a year
Used as a reference to validate the propagator
Comet C/2013 A1 (Siding Spring)
Propagation of the trajectory of the comet Siding Spring
Similar trajectory as the one of Snoopy [Farnocchia 2016]
Position available on SPICE for a year
Comet C/2013 A1 (Siding Spring)
Propagation of the trajectory of the comet Siding Spring Similar trajectory as the one of Snoopy [Farnocchia 2016]
Figure:Trajectory of Siding Spring (NASA)
Position available on SPICE for a year
Used as a reference to validate the propagator
Comet C/2013 A1 (Siding Spring)
Propagation of the trajectory of the comet Siding Spring Similar trajectory as the one of Snoopy [Farnocchia 2016]
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Figure:Trajectory of 2020 SO centered on the Earth (JPL)
Similar case study as Snoopy’s
Pre-Apollo era spacecraft Lost in heliocentric orbit
Long term propagation (54 years)
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Similar case study as Snoopy’s
Pre-Apollo era spacecraft Lost in heliocentric orbit
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Figure:Trajectory of 2020 SO centered on the Earth (JPL)
Similar case study as Snoopy’s
Pre-Apollo era spacecraft Lost in heliocentric orbit
Long term propagation (54 years)
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Figure:Trajectory of 2020 SO centered on the Earth (JPL)
Similar case study as Snoopy’s
Lost in heliocentric orbit
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Figure:Trajectory of 2020 SO centered on the Earth (JPL)
Similar case study as Snoopy’s Pre-Apollo era spacecraft
Lost in heliocentric orbit
Long term propagation (54 years)
2020 SO
Near Earth Object captured in October 2020 [Talbert 2020]
Identified as the Centaur rocket booster of Surveyor 2 launched on 20th September of 1966 [Nieberding 1968]
Figure:Trajectory of 2020 SO centered on the Earth (JPL)
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F
Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020] Snoopy’s trajectory and WT1190F’s seem close
No reentry observed for Snoopy
Has Snoopy reentered the Earth’s Hill Sphere ? (SOI)
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F
Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020] Snoopy’s trajectory and WT1190F’s seem close
No reentry observed for Snoopy
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F
Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020]
Snoopy’s trajectory and WT1190F’s seem close No reentry observed for Snoopy
Has Snoopy reentered the Earth’s Hill Sphere ? (SOI)
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020]
Snoopy’s trajectory and WT1190F’s seem close
No reentry observed for Snoopy
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020] Snoopy’s trajectory and WT1190F’s seem close
No reentry observed for Snoopy
Has Snoopy reentered the Earth’s Hill Sphere ? (SOI)
Snoopy
Apollo 10’s Lunar Module, jettisoned in 1969 in heliocentric orbit [Adamo 2012]
Figure:Apollo 10 LM Snoopy (NASA)
May have reentered in 2015 under the name WT1190F Investigated at ISAE-SUPAERO along with CNES
[Hautesserres 2020, Vagnone 2018, Villanueva Rourera 2020] Snoopy’s trajectory and WT1190F’s seem close
Monte-Carlo : The case of Snoopy
The aim is to estimatepR(t)the probability of Snoopy’s presence in
a sphere S(R) centered on the Earth, of radius R at a date t
The integral to estimate is then: pR(t) =
Z
R3
1S(R) kxkEarth f (x)dx (12)
With f the distribution function of Xi [Robert 2004]
Monte-Carlo estimator of pR(t) ˆ pR(t) = 1 N N X i=1 1S(R) kXikEarth (13)
Multiples of the radius RSOI =10−2AU of the SOI are
Monte-Carlo : The case of Snoopy
The aim is to estimate pR(t) the probability of Snoopy’s presence in
a sphere S(R) centered on the Earth, of radius R at a date t The integral to estimate is then:
pR(t) =
Z
R3
1S(R) kxkEarth f (x)dx (12)
With f the distribution function of Xi [Robert 2004]
Monte-Carlo estimator of pR(t) ˆ pR(t) = 1 N N X i=1 1S(R) kXikEarth (13)
Multiples of the radius RSOI =10−2AU of the SOI are
considered : Rn=nRSOI
Monte-Carlo : The case of Snoopy
The aim is to estimate pR(t) the probability of Snoopy’s presence in
a sphere S(R) centered on the Earth, of radius R at a date t The integral to estimate is then:
pR(t) =
Z
R3
1S(R) kxkEarth f (x)dx (12)
With f the distribution function of Xi [Robert 2004] Monte-Carlo estimator of pR(t) ˆ pR(t) = 1 N N X 1S(R) kXikEarth (13)
Multiples of the radius RSOI =10−2AU of the SOI are
Monte-Carlo : The case of Snoopy
The aim is to estimate pR(t) the probability of Snoopy’s presence in
a sphere S(R) centered on the Earth, of radius R at a date t The integral to estimate is then:
pR(t) =
Z
R3
1S(R) kxkEarth f (x)dx (12)
With f the distribution function of Xi [Robert 2004] Monte-Carlo estimator of pR(t) ˆ pR(t) = 1 N N X i=1 1S(R) kXikEarth (13)
Multiples of the radius RSOI =10−2AU of the SOI are
considered : Rn=nRSOI
Monte-Carlo : Error estimation
Since 1S(R) kXikEarth follows a Bernoulli law of parameter pR :
Var (1S(R) kXikEarth) = pR· (1 − pR)
If ∆ˆpR =3.09
√
ˆ
pR√·(1−ˆpR)
N , the central limit theorem ensures that :
ˆpR− ∆ˆpR, ˆpR+ ∆ˆpR
(14) is an estimation of the confidence interval at 99.9%, if Var (pR) 6=0
Since ˆpR· (1 − ˆpR) ≤ 14, the number of trajectories necessary to
estimate the probability can bounded by : ∆ˆpR≤ 3.09 1 2√N < 1.55 √ N (15) For instance N = 2.5.104 =⇒ ∆ˆpR <1%
Monte-Carlo : Error estimation
Since 1S(R) kXikEarth follows a Bernoulli law of parameter pR :
Var (1S(R) kXikEarth) = pR· (1 − pR)
If ∆ˆpR =3.09
√
ˆ
pR√·(1−ˆpR)
N , the central limit theorem ensures that :
ˆpR− ∆ˆpR, ˆpR+ ∆ˆpR
(14) is an estimation of the confidence interval at 99.9%, if Var (pR) 6=0
Since ˆpR· (1 − ˆpR) ≤ 14, the number of trajectories necessary to
estimate the probability can bounded by : ∆ˆpR≤ 3.09 1 2√N < 1.55 √ N (15) For instance N = 2.5.104 =⇒ ∆ˆpR <1%
Monte-Carlo : Error estimation
Since 1S(R) kXikEarth follows a Bernoulli law of parameter pR :
Var (1S(R) kXikEarth) = pR· (1 − pR)
If ∆ˆpR =3.09
√
ˆ
pR√·(1−ˆpR)
N , the central limit theorem ensures that :
ˆpR− ∆ˆpR, ˆpR+ ∆ˆpR
(14) is an estimation of the confidence interval at 99.9%, if Var (pR) 6=0
Since ˆpR· (1 − ˆpR) ≤ 14, the number of trajectories necessary to
estimate the probability can bounded by :
1 1.55
Monte-Carlo : Error estimation
Since 1S(R) kXikEarth follows a Bernoulli law of parameter pR :
Var (1S(R) kXikEarth) = pR· (1 − pR)
If ∆ˆpR =3.09
√
ˆ
pR√·(1−ˆpR)
N , the central limit theorem ensures that :
ˆpR− ∆ˆpR, ˆpR+ ∆ˆpR
(14) is an estimation of the confidence interval at 99.9%, if Var (pR) 6=0
Since ˆpR· (1 − ˆpR) ≤ 14, the number of trajectories necessary to
estimate the probability can bounded by : ∆ˆpR≤ 3.09 1 2√N < 1.55 √ N (15) For instance N = 2.5.104 =⇒ ∆ˆpR <1%
Validation methodology
Python prototype C++ program Dynamics validation Domain splitting validation Probability estimation validation Has Snoopy reentered in the Earth’s SOI ? Siding Spring 2020 SOFigure:Diagram of the validation methodology
Propagator validation : Siding Spring I
Positions normalized by AU = 1.5.108km, velocities by V =qμsun
AU :
Figure:Normalised error wrt Spice
Propagator validation : Siding Spring II
Propagator validation : Siding Spring III
Impact of domain splitting on the approximation error, at order 3 :
Figure:Evolution of the approximation error
2020 SO’s trajectory
Figure:2020 SO’s trajectory
2020 SO’s trajectory
TDA order : 3
Elapsed time : 115s Step : 1 day
2020 SO’s probability of presence
(a)Time dependence
TDA order : 5 Nb of domains : 8 Size sample : 25K Elapsed time : 58min TDA tolerance : 1e-11 Max TDA error : 1e-10 Step : 1 day
(b)Max probability
Figure:2020 SO’s probability of presence near Earth
Surveyor 2
TDA order : 5 Nb of domains : 16 Size sample : 100 Elapsed time : 100s Step : 1 daySnoopy’s trajectory
Figure:Snoopy’s trajectory
Snoopy’s trajectory
TDA order : 3 Elapsed time : 99s Step : 1 day
Figure:Snoopy’s trajectory
Snoopy’s probability of presence
TDA order : 5 Nb of domains : 16 Size sample : 25K Elapsed time : 70min TDA tolerance : 1e-11 Max TDA error : 1e-7 Step : 1 day
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error
TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max
error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max
error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error
TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max
error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error
TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max
error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Approximation error and speed-up
A set of 1000 trajectories of Snoopy were propagated to quantify the error made by TDA-based MC
Order of magnitude of the maximum relative error : 10−7 =⇒ Max
error ≈ 100km
This error does not affect the results on pR(t)
Approximation error close to the modelling and integration error TDA-based Monte-Carlo : 20 times faster than classic Monte-Carlo on Snoopy
Similar precision as classic Monte-Carlo
The code is still fully sequential, domain splitting may be done in parallel
Snoopy’s return
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Next close encounter with Earth : Between 2025 and 2027
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples
Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples
Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Next close encounter with Earth : Between 2025 and 2027
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Next close encounter with Earth : Between 2025 and 2027
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Next close encounter with Earth : Between 2025 and 2027
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ?
Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ? Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Next close encounter with Earth : Between 2025 and 2027
Conclusion
TDA propagator
Validated thanks to the comet Siding Spring
Faster than classic Monte-Carlo estimations for large samples Approximation error : Can be tuned so that it does not impact the probability estimations
Methodology validation
2020 SO is computed to be at least in S(R6)a few month after the
launch of Surveyor 2
It seems that Surveyor 2 is likely to be in S(R6)at that time
Did Snoopy reenter the Earth’s sphere of influence ? Close approach of Snoopy, at least in S(R7)
No statistical evidence that Snoopy entered the Earth’s SOI from 2000 to 2013
Future work
Parallelization of the computations
Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations
Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects
Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Parallelization of the computations Estimations on 2020 SO
Refining 2020 SO’s trajectory
Thorough propagation of Surveyor 2 launcher to compare it with 2020 SO
Estimations on Snoopy
If Snoopy did not reenter, what is WT1190F ?
Propagating WT1190F’s trajectory to compare to two objects
Propagation of Lunar Prospector Trans-lunar injection stage, other candidate for WT1190F
Future work
Figure:WT1190F’s preliminary trajectory
References I
Daniel R. Adamo.
Earth Departure Trajectory Reconstruction of Apollo Program Components Undergoing Disposal in Interplanetary Space.
http://www.aiaahouston.org, 2012.
R. Armellin, P. Di Lizia, F. Bernelli-Zazzera et M. Berz.
Asteroid Close Encounters Characterization Using Differential Algebra: The Case of Apophis.
Springer, 2010. Martin Berz.
Modern map methods in particle beam physics.
References II
D. Farnocchia, S. R. Chesley, M. Micheli, A. Delamere, R. S. Heyd, D. J. Tholen, J. D. Giorgini, W. M. Owen et L. K. Tamppari.
High Precision Comet Trajectory Estimates: the Mars Flyby of C/2013 A1 (Siding Spring).
Icarus, 2016.
William M. Folkner, James G. Williams, Dale H. Boggs, Ryan S. Park, et Petr Kuchynka.
The Planetary and Lunar Ephemerides DE430 and DE431.
Rapport technique, Jet Propulsion Laboratory, California Institute of Technology, 2014.
R. M. Georgevic.
Mathematical Model of the Solar Radiation Forceand Torques Acting on the Componentsof a Spacecraft.
Rapport technique, Jet Propulsion Laboratory, 1971.
References III
A. Hanley J A Lippman-Hand.
If Nothing Goes Wrong, Is Everything All Right? Interpreting Zero Numerators.
Journal of the American Medical Association, 1983.
D. Hautesserres, L. Villanueva Rourera et P. Guardabasso.
Research of the History of WT1190F and that of Snoopy.
Rapport technique, Centre National d’Etudes Spatiales (CNES) and Institut Supérieur de l’Aéronautique et de l’Espace
(ISAE-SUPAERO), 2020.
Dario Izzo et Francesco Biscani.
References IV
Mauro Massari, Pierluigi Di Lizia et Mirco Rasotto.
Nonlinear Uncertainty Propagation in Astrodynamics Using Differential Algebra and Graphics Processing Units.
American Institute of Aeronautics and Astronautics, 2017. John J. Nieberding.
Atlas-centaur Flight Performance for Surveyor Mission B, Appendix A.
Rapport technique, NASA, 1968. Christian P. Robert et George Casella.
Monte carlo statistical methods.
Springer, 2004.
References V
Tricia Talbert.
New Data Confirm 2020 SO to be the Upper Centaur Rocket Booster from the 1960’s.
https://www.nasa.gov/feature/new-data-confirm-2020-so-to-be-the-upper-centaur-rocket-booster-from-the-1960-s, 2020.
Accessed on the 30/01/2021.
Federica Vagnone.
Snoopy’s Trajectory – Debris Identification.
References VI
Lydia Villanueva Rourera, Stéphanie Lizy-Destrez et Paolo Guardabasso.
Snoopy’s Trajectory - Debris Identification.
Rapport technique, Institut Supérieur de l’Aéronautique et de l’Espace (ISAE-SUPAERO), 2020.
A. Wittig, P. Di Lizia, R. Armellin, K. Makino, F. Bernelli-Zazzera et M. Berz.
Propagation of Large Uncertainty Sets in Orbital Dynamics by Automatic Domain Splitting.
Springer, Celest Mech Dyn Astr, 2015.
Propagating uncertainties using TDA I
Example
Solving the following Cauchy problem numerically : ( ˙y(t) = f (y(t), t)
y (0) = y0
For real numbers, Euler’s explicit method can be used :
yn+1 =yn+h · f (yn,tn) +O(h2) (16)
If y0is tainted with an error
δ
y0 =⇒ an other propagation isPropagating uncertainties using TDA I
Example
Solving the following Cauchy problem numerically : ( ˙y(t) = f (y(t), t)
y (0) = y0
For real numbers, Euler’s explicit method can be used :
yn+1 =yn+h · f (yn,tn) +O(h2) (16)
If y0is tainted with an error
δ
y0 =⇒ an other propagation isneeded
Compared performances I
Figure:Uncertainty propagation using real numbers
Compared performances I
Compared performances I
Figure:Uncertainty propagation using real numbers
Compared performances I
Compared performances I
Figure:Uncertainty propagation using real numbers
Compared performances I
Compared performances I
Figure:Uncertainty propagation using real numbers
Compared performances I
Compared performances II
Figure:Uncertainty propagation using TDA
Compared performances II
Compared performances II
Figure:Uncertainty propagation using TDA
Compared performances II
Compared performances II
Figure:Uncertainty propagation using TDA
Compared performances II
Compared performances III
Computation time of N trajectories:
Real numbers: N propagations are needed:
∆treal =N∆tprop,real (17)
TDA: 1 propagation and N evaluations are needed:
∆tpoly = ∆tprop,poly+N∆teval (18)
The speed-up rN provided by TDA is:
rN = ∆tprop,poly N∆tprop,real + ∆teval ∆tprop,real → ∆teval ∆tprop,real (19)
Compared performances III
Computation time of N trajectories:
Real numbers: N propagations are needed:
∆treal =N∆tprop,real (17)
TDA: 1 propagation and N evaluations are needed:
∆tpoly = ∆tprop,poly+N∆teval (18)
The speed-up rN provided by TDA is:
rN = ∆tprop,poly N∆tprop,real + ∆teval ∆tprop,real → ∆teval ∆tprop,real (19)
Compared performances III
Computation time of N trajectories:
Real numbers: N propagations are needed:
∆treal =N∆tprop,real (17)
TDA: 1 propagation and N evaluations are needed:
∆tpoly = ∆tprop,poly+N∆teval (18)
The speed-up rN provided by TDA is:
rN = ∆tprop,poly N∆tprop,real + ∆teval ∆tprop,real → ∆teval ∆tprop,real (19)
Compared performances III
Computation time of N trajectories:
Real numbers: N propagations are needed:
∆treal =N∆tprop,real (17)
TDA: 1 propagation and N evaluations are needed:
∆tpoly = ∆tprop,poly+N∆teval (18)
The speed-up rN provided by TDA is:
rN = ∆tprop,poly N∆tprop,real + ∆teval ∆tprop,real → ∆teval ∆tprop,real (19)
Compared performances IV
Figure:Speed-up provided by TDA-based Monte-Carlo
Introduction to Monte-Carlo estimations I
The aim is to estimate the following integral: I =
Z
Rd
φ
(x )f (x )dx (20)
With f a density of probability, and d ∈ N the dimension of the problem. [Robert 2004]
The Monte-Carlo (MC) estimator of I is: ˆIN = 1 N N X i=1
φ
(Xi) (21)Where N ∈ N and the random variables in the sequence (Xi)i∈N
are independent and identically distributed (IID) random variables following the distribution f .
Introduction to Monte-Carlo estimations I
The aim is to estimate the following integral: I =
Z
Rd
φ
(x )f (x )dx (20)
With f a density of probability, and d ∈ N the dimension of the problem. [Robert 2004]
The Monte-Carlo (MC) estimator of I is: ˆIN = 1 N N X i=1
φ
(Xi) (21)Where N ∈ N and the random variables in the sequence (Xi)i∈N
are independent and identically distributed (IID) random variables following the distribution f .
Introduction to Monte-Carlo estimations II
Theorem (Convegence of the MC estimator)
If ˆIN is the MC estimator of I, then when N → ∞:
ˆIN −→ E(
φ
(X1)) =I (22)Monte-Carlo : Error estimation II
The confidence interval cannot by computed thanks to CLT if Var (pR) =0. Thus, if ˆpR ∈ {0, 1}
If ˆpR=0, the confidence interval at 99.9% is estimated as
[Hanley 1983] : ˆ pR∈ 0,6.9 N (23) If ˆpR=1, the confidence interval at 99.9% is estimated in the same
way : ˆ pR ∈ 1 − 6.9 N ,1 (24)