See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318143703

Efficient Fast Approximation for Aircraft Fuel Consumption for Decision-

Making and Policy Analysis

Conference Paper · June 2017

DOI: 10.2514/6.2017-3338

CITATIONS READS
2 878

2 authors:

Jefry Yanto Rhea P. Liem

The Hong Kong University of Science and Technology The Hong Kong University of Science and Technology
4 PUBLICATIONS   15 CITATIONS    21 PUBLICATIONS   216 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Amphibious Aircraft Design Framework View project

Aircraft fuel burn model for budgeting purpose View project

All content following this page was uploaded by Jefry Yanto on 29 October 2018.

The user has requested enhancement of the downloaded file.

 Efficient Fast Approximation for Aircraft Fuel Consumption
for Decision-Making and Policy Analysis

Jefry Yanto* , Rhea P. Liem†

Hong Kong University of Science and Technology (HKUST), Hong Kong

Environmental impacts of aviation and fuel economy have been the main drivers in aircraft design
and policy analysis in the aviation sector. The increasing demand in air transportation and the volatil-
ity of fuel price push for fuel reduction measures to be implemented. These measures might come
from changes in technology (e.g., improved aircraft design or engine), operations, or both. To make
decisions on the best policy to implement, several policy scenarios need to be carefully evaluated and
compared. This calls for a fast, efficient, and yet accurate fuel burn computation method, in partic-
ular to compute the total aggregate fuel burn given a set of flight missions. The available low-fidelity
fuel burn models are typically not accurate enough. The high-fidelity models, on the other hand, are
too computationally expensive and thus can quickly become intractable when we need to evaluate
hundreds of thousands of flight missions. We therefore aim to develop a surrogate model to approx-
imate the total aggregate fuel burn, where one model is derived for each aircraft type. By doing so,
the cost to compute the total aggregate fuel burn is a function of the number of aircraft types, instead
of the number of total flight missions; this will significantly reduce the computational burden. We
in particular choose regression model, which is a sample-based surrogate modeling technique, due to
its simplicity and non-intrusive nature. To generate the fuel burn database, we develop a medium-
fidelity, data-enhanced fuel burn approximation model. We use data obtained from running trajec-
tory simulations (from EUROCONTROL’s BADA) to derive some correction factors to complement
the classical Breguet range equation, which is commonly used to evaluate aircraft performance. This
fuel burn database is generated using actual flight mission data obtained from the Bureau of Trans-
portation Statistics (BTS), corresponding to 37 aircraft types and around 286 thousands of flights.
The validation shows that the percentage root mean square deviation (RMSD) errors for the total ag-
gregate fuel burn computation are below 1% for all aircraft types, which are deemed very accurate.
The derived regression models provide the sensitivity information of the fuel burn performance. This
information can then be used in a clustering algorithm, to identify aircraft types that have similar
fuel burn performance and group them together. With the formed clusters, we can further reduce the
problem dimensions of the total aggregate fuel burn computation, as it now becomes a function of the
number of clusters.

I. Introduction
The advancement of numerical simulation has assisted considerably the study of many complex physical phenom-
ena and is becoming increasingly widespread as a means to support decision-making and policy making processes.
Such computational models are often complex, involving many disciplines, many input parameters, and long com-
putation times. The airline industry has relied a lot on such numerical analyses and modeling in its decision making
analyses. Regulatory agencies analyze and compare a number of policy scenarios prior to enforcing a new policy
or regulation. The modeling of the air transportation system encompasses different aspects such as pure science and
engineering (to model the aircraft system and its operation), economic (for the cost analysis), environmental impact,
etc. It is thus imperative to take a system architecture approach that takes into account the various interactions and
information/data flow from one subsystem to another. For instance, to support aviation environmental policy-making,
a computational model is required to evaluate the impacts of environmental policy alternatives in the form of social
costs, such as public and private mitigation costs and public environmental benefits [1].
These evaluations require outputs from flight performance analyses, in particular the amount of fuel burned during
any given flight mission (will be simply referred to as fuel burn for the remaining of this paper). With the expected
steady increases in the demand of air transportation [2] and the volatility of fuel prices [3], fuel economy and environ-
mental impacts of aviation have become the main drivers in many airline and air traffic policy and decision making
processes. In 2010, international aviation consumed approximately 142 million metric tonnes of fuel, resulting in 448
million metric tonnes (Mt, 1kg x 108 ) of CO2 emissions and the fuel consumption is projected to multiply by 2.8 to 2.9
* M.Phil. Candidate, AIAA Student Member
† Assistant Professor, AIAA Member

1 of 22

American Institute of Aeronautics and Astronautics

times by 2040 [4]. The direct operating cost (DOC) directly depends on the fuel burn. In addition, fuel consumption
is a direct surrogate for CO2 emissions, which is an important greenhouse gas, and also an indirect indicator for other
aircraft emissions, such as NOx and H2 O [5]. Aircraft emissions also contain components that are the predominant
air pollutants causing health damages. Some of the health impacts include: premature mortality for adults and infants,
chronic bronchitis, respiratory problems, cardiovascular problems, and asthma [6].
The scale and complexity of the required fuel burn analyses when considering the global air transportation system
are immense. The simulation of all flights within one year involves over 35 million flights with approximately 350
aircraft types and thousands of input parameters [7], analyzed with models spanning airline economics, environmen-
tal economics, aircraft operations, aircraft performance and emissions, noise, local air quality, and global climate.
Different analyses and evaluations might have different scenarios. To evaluate local/regional environmental impacts,
for instance, the analysis would require the total fuel burn at a certain time to/from a certain airport. On the other
hand, for the operating cost evaluation of an airline it might make more sense to aggregate the total fuel burn for the
same airline, or aircraft types, etc. To reduce the global fuel consumption, policy makers might consider different
policy scenarios about the future. These scenarios might involve the introduction of new tachnology and/or change in
aircraft operations. This might include an introduction of new aircraft engines, new airframe designs, new routes, or
any of their combinations. Ideally airliner would want to assign the most efficient aircraft for each flight mission. For
instance, a direct flight of 8 000 nautical miles might burn more than twice the fuel of a 4 000 nautical miles flight [8].
Such an observation might change an airline’s flight route distribution, e.g., the cancellation of Singapore Airlines’
direct flight from New York to Singapore in 2013 [9]. To answer those questions, we need to have a simple, efficient,
and yet reliable fuel burn approximation method. Moreover, the prediction of the total aggregate fuel burn would be
more relevant to what policy makers required.
There have been a wide variety of fuel-burn approximation models spanning from the low-fidelity to high-fidelity
models. There has always been a tradeoff between the accuracy and efficiency in any computational analyses. High-
fidelity models for fuel burn computation perform accurately, however, it is typically computationally expensive and
time consuming [10, 11]. Low-fidelity models can reduce the computational time, but at the expense of the accu-
racy [12]. It is common to assume a cruise-only profile in computing the fuel burn for a flight mission [13, 14]. While
the cruise fuel burn might be dominant in long-haul flights, for short-haul flights the fuel-burn contribution from the
climb segment might be too significant to be ignored [15].
The main objective of this research is to develop an efficient fast approximation model for aggregate fuel burn
computation, and yet accurate. A simple sample-based surrogate is derived for each aircraft type to facilitate this
computation. For this purpose, we first need to generate a fuel-burn database for each aircraft type. To generate
the database, we consult the US flight mission data corresponding to 2015 from a publicly available database. To
compute the fuel burn for each flight mission, we develop a medium fidelity fuel burn computation method, where
we incorporate data-driven knowledge (from running trajectory simulations) into the low-fidelity model to improve its
accuracy, while maintaining the computational efficiency. The trajectory simulation results are also used to identify the
“right” trajectory for short-haul flights. We often find that the computed climb distance is longer than the total mission
range of a short-haul flight, which typically leads to the cruise-only assumption in the modeling. We found that this
assumption would overestimate the fuel burn, which would lead to a large approximation error. The derived surrogate
model would allow us to efficiently compute the total fuel burn for an aircraft type. When there are a large number
of aircraft types to analyze, we explore a way to further reduce the problem dimension by performing a clustering
algroithm, which could group aircraft with similar fuel burn performance together. In which case, the computation of
total aggregate fuel burn would be a function of the number of clusters, instead of the number of aircraft types. Each
method developed in this work is properly validated.
This paper starts with a brief fuel burn computation review in Section II. We will then present our proposed
approach in Section III and describe the developed fuel burn computation method in Section IV. Results, including
their validations, will be explained in Section V. We conclude this paper with a brief summary, conclusion, and future
work in Section VI.

II. Fuel Burn Computation Review

Figure 1 illustrates the “ideal” fuel burn evaluation framework. Assessing aircraft fuel burn is complex [16],
since we need to simultaneously consider the engine and airframe designs, aircraft operation, atmospheric quantities
(ambient temperature, pressure, density, and speed of sound), and airline economics. To obtain global fuel burn
and emissions inventories, we need to simulate the flights of all aircraft flown worldwide, which requires a massive
amount of computation. We first need to obtain the detailed fleet descriptions and flight schedules [17], to derive the
appropriate list of mission parameters to simulate.
Due to the aforementioned complexities and immense computational cost, aircraft fuel-burn computations are

2 of 22

American Institute of Aeronautics and Astronautics

 Aircraft
geometry
Climb
Cruise Descent Fuel burn,
emissions
Detailed mission analysis procedure
Mission Takeoff Landing
parameters • Aircraft operating
cost
• Environmental
Engine Engine model Atmospheric Weight-and-balance impacts
parameters model model
• Policy analysis

Atmospheric Aircraft components

conditions and weights

Figure 1: The disciplines involved and complexities in computing aircraft fuel burn.

typically done with a simplification of either the physics in the model or the mission profile considered [10, 18].
Analytical and empirical models are sometimes used to reduce the computational time [19], at the expense of accuracy.
Most fuel burn computations are derived based on the classical Breguet range equation [12, 20, 21, 22], which is
expressed as [23, 24]:  
V L Wi
R= ln , (1)
cT D Wf
where R, V , and L/D refer to the mission range, flight speed, aerodynamic lift-to-drag ratio, respectively. The variable
cT denotes the thrust specific fuel consumption (TSFC), which is defined as the weight of fuel burned per unit time
divided by the unit thrust. TSFC is a property of the aircraft engine, which is assumed constant in this work. Wi and
Wf refer to the initial and final segment’s aircraft weight, respectively. This equation, however, is only applicable
under the assumption that TSFC, L/D, and flight speed are constant. One important implication, however, is that the
takeoff, climb, and descent segments are not properly modeled by this equation [25].
It is common to compute the fuel burn of a flight mission based only on the cruise segment, omitting the climb and
descent segments [13, 14]. In this practice, the total mission range becomes the cruise range. Dancila [26] performed
fuel burn rate time-dependent model on cruise segment. This computation splits the cruise segment into sub-segments
to obtain the gross weight via an iteration procedure. This approach was claimed to have more accurate perfomance
than the Breguet range equation.
When other segments are included, their fuel burns are typically computed using the appropriate fuel fractions.
Roskam [27] defined the fuel fraction as the ratio between final weight and initial weight on each segment. The fuel
fraction values for engine start and warm-up, taxi, take-off, climb, descent, and landing until shutdown are typically
0.99, 0.99, 0.99, 0.985, 0.99, 0.99, respectively. Then the fuel required for each segment can be calculated using the
fuel fraction and the segment’s initial and final weights. O’Kelly [28] modeled the fuel burn for taxi, takeoff, cruise,
and landing with distance and aircraft size as inputs. This model, however, does not model the fuel burn for climb and
descent segments. These approaches are simple and computationally inexpensive but the result might not be accurate.
For a more detailed and realistic fuel burn computation that takes into account other segments (e.g., climb and
descent), we can employ either higher-fidelity models or run trajectory simulations. These methods can take into
account the different flight conditions along the mission profile, and thus offer more realistic results. However, they
are computationally expensive and take a long time to run. Completing the run for one flight mission with a high-
fidelity model might take minutes, and might need to be done in parallel. While this might be acceptable to analyze a
few flight missions, the required computational time and cost can quickly become intractable as we include thousands
or millions of missions.
For the flight trajectory simulations, EUROCONTROLa has created and maintained an aircraft performance model
(APM) called the Base of Aircraft Data (BADA)b . BADA provides aircraft performance and operations models to
support trajectory simulation for various aircraft types [29, 30] and is widely used especially in Air Traffic Management
(ATM) applications. To run BADA simulations, however, requires a license issued by EUROCONTROL. Moreover,
for each flight simulation, the user needs to input the aircraft and flight information one by one (e.g., aircraft type,
initial and final altitudes during climb) for each segment. Dalmau and Prats [31] argued that the derivation of fuel
consumption in BADA, in particular the fuel approximation at the terminal area, is not accurate enough. While each
a https://www.eurocontrol.int/
b http://www.eurocontrol.int/services/bada

3 of 22

American Institute of Aeronautics and Astronautics

simulation only takes a few seconds to run, it is not very practical to run the trajectory simulations for the entire
mission profile (including climb, cruise, and descent) thousands of time manually to generate a fuel burn database.
These runs, however, would be used to derive some performance parameters in our fuel burn computation method, as
will be further discussed later.
Several efforts have been done by various researchers to include the climb and descent segments in estimating
the mission’s fuel burn without running trajectory simulations. Lee and Chatterji [32], for instance, further improved
the fuel burn approximations for climb and descent segments by applying a correction factor. The climb segment’s
fuel burn was computed by applying a correction factor to the amount of fuel required to cruise at the same distance;
this correction value was represented by a fraction of the takeoff weight. The fuel burn for descent segment was
approximated with the fuel burn required to cruise the ground distance from top-of-descent to landing airport. The
fuel burn for the cruise segment was computed based on the Breguet range equation, and the fuel burn for taxi, takeoff,
approach and landing could be collectively approximated as a fraction of takeoff weight. A factor of 0.007 was used
for the total of those phases, following the work by Kroo [33]. The reserve fuel was expressed in fuel fraction;
generally a factor of 0.08 of zero-fuel weight [32, 33]. This approach, albeit simple, improved the accuracy of the fuel
prediction for the climb and descent segments. However, we observed some discrepancies in the climb and descent
fuel approximations upon comparing the results to the fuel burn obtained from running flight trajectory simulations on
BADA, as will be discussed in Section V.
For a higher-fidelity mission analysis modeling, Liem [18] performed a detailed mission analysis procedure that
modeled different flight conditions for the climb, cruise, and descent segments. Fuel fraction method is used to
approximate the fuel burn during startup, taxi, takeoff, and landing. Solving the governing equation of each mission
segment (climb, cruise, and descent) via numerical integration overcomes the limitation of low-fidelity models. This
technique required the aerodynamic performance information (lift, drag, and moment coefficients) at different flight
conditions along the mission profile to perform the numerical integration. With the numerical integration procedure
and its iterative nature, this mission analysis procedure is computationally expensive and might take hours to complete
with 16 processors. To reduce the required computational time, surrogate models were employed to approximate the
aerodynamic performance coefficients required in the analyses. Using surrogate models offered a speedup of at least
70 times [11], such that each flight mission analysis was completed in 6–18 minutes (depending on the mission profile)
with 16 processors. Let say we need to compute the total fuel burn of 10 000 flight missions. Using the same number
of processors, this computation would take around 40 days to complete, which is impractical.

III. Proposed Approach

This section provides an overview of the proposed framework. The step-by-step procedure is presented, and each
step will be further discussed in the next section.
The main objective of this research is to derive a fast method to approximate the total fuel burn for any given
sets of flight missions. In particular, we derive a surrogate model for each aircraft type. We employ a sample-based
surrogate modeling technique to derive the fuel burn approximation model, due to its simplicity and non-intrusive
nature. Such a technique requires a large number of samples to achieve the desired level of accuracy. For this purpose,
we first need to have a fuel burn database for the sample population. A medium-fidelity, data-enhanced fuel burn
approximation model is derived for this purpose, which will be described in details in the next section. Having derived
the surrogate model for each aircraft type, the computational cost of approximating the total fuel burn of a given set
of flight missions is thus a function of the number of aircraft types included in the set. We then argue that we could
further reduce the problem dimension by grouping similar aircraft types together. This grouping exercise could be
done by employing an unsupervised learning algorithm, which will also be demonstrated in this work. This procedure
is summarized as follows:

Step 1 Obtain the list of flight missions to be analyzed (flight mission inventories).
For this purpose, we select a number of aircraft types, and obtain the mission information of flights to and from
a specific region. We used the publicly available data from the United States Department of Transportation’s
Bureau of Transportation Statistics (BTS)c , corresponding to all flights to and from the United States in 2015.
For each aircraft type, we gather information on the relevant parameters (which will be further discussed in
Section V), mainly from the aircraft manufacturers’ websites.

Step 2 Generate the fuel burn inventories by simulating each flight mission individually.
Each flight mission is divided into five segments (takeoff, climb, cruise, descent, and landing), and the fuel burn
for each segment is estimated separately. Figure 2 shows a simple illustration of the general mission profile
c TranStats, Bureau of Transportation Statistics http://www.transtats.bts.gov/

4 of 22

American Institute of Aeronautics and Astronautics

 considered in this work. The total fuel burn for each flight mission would include those of the segments shown
in Figure 2 and the maneuver fuel burn. The detailed procedure to compute flight mission’s fuel burn will be
given in Section IV.

W0 Cruise W1

Descent
Climb

Takeoff WTO WZF Landing

Figure 2: Simplified flight mission with five segments: takeoff, climb, cruise, descent, and landing.

Step 3 Derive the surrogate model for each aircraft type.

Surrogate modeling techniques have been commonly used to reduce the cost and time of the computational
procedure by providing low-cost substitutes to replace expensive evaluations of the original physics-based mod-
els [34]. A surrogate model uses mathematical models to provide a simpler approximation of a physical system,
thereby reducing the computational expenses of analysis and optimization [35, 36]. Surrogate modeling tech-
niques span from simple models such as polynomial regressions [37] and interpolations [18]. Regression models
are derived in a least-squares sense and more suitable for functions with inherent random error while interpola-
tions are more appropiate to model a function with no measurement (random) error component is involved [18].
Regression models are selected in this work for the simplicity and the derived expression can also give insights
into the sensitivity of the output with respect to the input variables. Using the fuel burn database corresponding
to the BTS mission data, we derive a two-dimensional regression model to approximate fuel burn for each air-
craft type, using the payload and range as input variables. These two variables are selected as the computed fuel
burns show linear relationships with them, as shown in Figure 3. The plots are only shown for two representative
aircraft, but similar trends are observed in other aircraft types as well.

(a) Boeing B737-400 (b) Canadair RJ200

Figure 3: Regression models for fuel consumption with range and payload as input variables

Figure 4 illustrates the Steps 1, 2, and 3 described above. This procedure is repeated for each aircraft type
included in the analysis.
Step 4 Demonstrate the use of clustering algorithm to group similar aircraft together
The gradient information provided by the regression models gives insight into the sensitivity of fuel burn with
respect to the mission payload and range. This sensitivity information can then be used to study any similar-
ity patterns among the different aircraft types, by employing some clustering algorithms. By identifying and

5 of 22

American Institute of Aeronautics and Astronautics

 Flight
mission Fuel burn Fuel burn Regression
database computation database model
(BTS)

Figure 4: Fuel burn surrogate model is derived for each aircraft type, based on the fuel burn database generated for the BTS mission data.

grouping aircraft with similar fuel burn performance together, we hope to reduce the computation required to
approximate the total aggregate fuel burn.

IV. Fuel Burn Approximation Method

This section describes the procedure to derive the fuel burn approximation model. We first describe the flight
mission profiles adopted in this work. We use different mission profiles for short- and long-haul flights, for reasons
that would be explained shortly. We then present the procedure to compute fuel burn of a flight mission, by modeling
each segment separately. A regression model is then derived for each aircraft type, followed by the clustering algorithm
to group similar aircraft types together.
Mission profile
In this work, we consider different mission profiles depending on the mission range (short-haul and long-haul mis-
sions). For the long-haul missions, the flight mission profile is divided into five segments as shown in Figure 2. This
mission profile has been commonly used in many flight mission analysis procedures [10, 32]. We found, however, that
this mission profile is not suitable for short-haul missions, as will be discussed shortly.
In most flights, the cruise segment is considered as the most significant flight segment. In fact, it is common to
use the cruise fuel burn as a surrogate of the entire flight mission’s fuel burn [14]. We initially employ this approach
to compute the fuel burn of some short-haul flights, i.e., when the climb and descent segments could not be modeled
properly due to the short mission ranges. However, we observed some peculiar results as shown for three representative
aircraft types in Figure 5. In these plots, the computed fuel burn corresponding to the short-haul flights are highlighted
within the red ellipses. The rest of data points shown in this figure were computed by considering all five flight
segments (Figure 2) and maneuver segment. This observation suggests that for short-haul flights, using the cruise fuel
burn to represent the flight mission’s fuel burn would overestimate the fuel burn computation. This inaccuracy would
deem any subsequent analysis results irrelevant. We therefore need to model the flight mission profile differently for
short-haul flights, to give more realistic results.
Fuel burn (kg)

Fuel burn (kg)

Fuel burn (kg)

Payload (kg)

Payload (kg)

Payload (kg)
Short-haul

Short-haul Short-haul

Distance (km) Distance (km) Distance (km)

(a) Airbus A321 (b) Boeing B737-900 (c) Canadair RJ700

Figure 5: Fuel burns for all flight missions corresponding to three representative aircraft types. Since the five-segment mission profile assumption cannot be implemented
to short-haul flights, their fuel burn computations are performed by assuming a cruise-only profile. The results suggest that this assumption overestimates the
actual fuel burn.

Based on this finding, we model the short-haul mission profile by assuming a saw-tooth profile. This profile
assumes that the flight would not reach the cruise segment; instead, it goes straight to descent upon climbing [38].
The ratio of ground distance for climb and descent with this approach is equivalent to ratio of ground distance for
climb and descent for long-haul mission, which can be proven using the triangle similarity theorem. This profile has
previously been used by Simos and Jenkinson [15]. Figure 6 shows the saw-tooth mission profile for the short-haul

6 of 22

American Institute of Aeronautics and Astronautics

flights. We classify a flight mission into short-haul flight if the total flight mission range is less than the total of ground
range of climb and descent segments, and likewise we classify into long-haul if the total flight mission range is more
than total range of climb and descent segment.

Top of climb

Climb Descent

Takeoff WTO WZF Landing

Figure 6: Simplified flight mission profile (the saw-tooth profile) for short-haul mission.

Segment-by-segment fuel burn computation procedure

The proposed segment-by-segment fuel burn computation procedure is illustrated in Figure 7. As inputs we have
engine parameter, aircraft performance parameters, and mission parameters. As shown in the figure, we use BADA
trajectory simulation results in the procedure, which will be described shortly. The output of this procedure is the fuel
burn of the flight mission.

Engine
Parameters TSFC
Cruise

L/D ratio
Cruise altitude Descent
Aircraft
Cruise mach
Performance
number
Climb Fuel burn
Parameters
Climb & descent
speed

Takeoff Landing
Mission Range
Parameters Payload

Fuel increment & Breguet

Climb &
fuel decrement range
descent range
factor model

Segment’s Fuel factor

range model model

BADA Flight Trajectory

Simulation

Figure 7: Proposed segment-by-segment fuel burn computation framework.

We use BADA trajectory simulation results to derive the climb and descent distances, as well as their corresponding
fuel correction factors for each aircraft type. The number of flight mission samples required to derive the climb and
descent fuel correction factors is determined based on some statistical measures, which will be further discussed in
Section V of this paper. We use 10 samples for each aircraft type, which is manageable to run on BADA simulations.
To compute the fuel burn required for each flight mission, we consider simple mission profiles as shown in Fig-
ure 2 for long-haul flights and Figure 6 for short-haul flights. Although an aircraft typically climbs and descends
through a few subsegments (i.e., accelerated climb/decelerated descend, constant velocity climb/descent, constant
Mach climb/descent) [10], at this stage we combine those subsegments into one climb or one descent segment. In
addition to the fuel burn during climb (Wf, climb ), cruise (Wf, cruise ), and descent segments (Wf, descent ), the total fuel
burn computation also includes the maneuver (Wf, man ) and reserve fuels (Wf, res ). The maneuver fuel is combination
of warm-up, taxi, takeoff, approach, and landing phases [39]. The total fuel burn would then obtained from summing
up all these components, as shown below:

Wf = Wf, climb + Wf, cruise + Wf, descent + Wf, man + Wf, res (2)

7 of 22

American Institute of Aeronautics and Astronautics

 The procedure to compute each fuel burn component is described below. Note that the procedure is not exactly
sequential (i.e., from the first mission segment up to the last one). This is done because the fuel burn computation
is interrelated, where one computation might need information from another. This procedure is also non-iterative, as
shown in the flowchart presented in Figure 8.

Step 1 Compute the range (ground distance) traveled during descent, Rdescent . As shown in Figure 2, we assume the
descent segment to start from the cruise altitude and end at 0 altitude. The descent range was determined by first
observing BADA flight trajectory simulation results for all flight missions considered in this study, as shown in
Figure 9. Each dot corresponds to a single flight mission, whereas each color corresponds to a unique aircraft
type. This plot shows how the descent distance varies with mission payload. From this plot we can observe that
some aircraft types exhibit linear relations between the descent distance and payload, whereas for other aircraft
types the relations are nearly constant. Based on this observation, we differentiate two categories of aircraft
types based on their descent range-and-payload relationship. The K-means clustering algorithm is employed
to divide the aircraft type into two groups. From the results, it turns out that the bigger aircraft (e.g. Airbus
A340, Boeing B777) tend to have a constant descent range, whereas the smaller ones (e.g Canadair RJ-700,
Embraer 140) have stronger linear relationships. For the latter group, the linear relationship is derived based on
these BADA results. Each group is validated with standard deviation hypothesis test to check whether the range
sample point for each aircraft type can be assume as constant value. This range categorization, however, is only
applicable for long-haul mission flights. We observed that the ratio between climb and descent ranges varies
linearly with respect to the payload weight. We thus use this ratio to determine the climb and descent ranges for
short-haul aircraft.
Step 2 Compute the range (ground distance) traveled during climb, Rclimb . We assume the climb segment to start from
the 0 altitude and end at cruise altitude as shown in Figure 2. The climb ranges for 10 samples corresponding
to each aircraft type as obtained from the BADA flight trajectory simulations are shown in Figure 10. From
these simulation data we derived the regression model for climb range for each aircraft type. However, this
model is only applicable when the flight mission is categorized as long-haul mission. When the flight mission
is categorized as short-haul mission, the climb range is calculated from ratio of climb-descent. Before we can
compute the climb fuel burn, we first need to find the aircraft weight at the end of climb segment, which we can
easily obtain once we compute the cruise fuel burn, which is done in the next step.
Step 3 Compute Wf, descent , by assuming that the amount of fuel burned during descent equals to fuel burned during
cruise for the same flight range (W̃f, cruise ) minus the descent fuel factor as:

Wf, descent = W̃f,eqcruise − fdescent WZF + C. (3)

W̃f,eqcruise is obtained using the Breguet range equation. fdescent and C are derived based on BADA data (as
correction factor) and WZF refers to the zero-fuel weight.
As mentioned in Section II, Lee and Chatterji [32] approximated the descent segment by assuming the fuel re-
quired is equivalent to amount of fuel required to cruise the same flight distance. Upon comparing this approach
to the BADA flight trajectory simulation results, however, we observed some discrepancies. In particular, the
use of the Breguet range equation directly for the descent segment overestimated the amount of fuel burn, which
are shown in Figure 11. With the derived correction factor, we can obtain a more accurate fuel burn estimate for
the descent segment, using the BADA simulation results as the reference.
Step 4 Compute the fuel burn during cruise segment, Wcruise . Since the climb and descent ranges have been obtained,
we can first compute the cruise range, Rcruise , as follows:

Rcruise = R − Rclimb − Rdescent − Rtakeoff − Rlanding (4)

Rtakeoff and Rlanding referred to distance required for takeoff and landing, which are obtained from website
SKYbraryd . We assume that Rtakeoff and Rlanding are constant for each aircraft type. The cruise fuel burn,
Wcruise = W0 − W1 , can then be found by implementing the Breguet range equation (Eqn. 1), with W0 and W1
as the initial and final weight of the cruise segment, respectively.
d http://www.skybrary.aero/index.php/Category:Aircraft

8 of 22

American Institute of Aeronautics and Astronautics

 Start

Calculate
descent’s
range

Calculate
climb’s
range

Categorized Climb-
as short-haul descent
mission? Yes ratio

No
Update
Calculate
climb and
cruise’s
descent’s
range
range

Calculate
descent’s Calculate
fuel descent’s
fuel
Calculate
cruise’s fuel
Calculate
climb’s fuel
Calculate
climb’s fuel

Calculate
reserve and
maneuver’s
fuel

Calculate
total fuel
burn

End

Figure 8: Fuel burn computation procedure.

9 of 22

American Institute of Aeronautics and Astronautics

Figure 9: Relation between the ground distance for descent segment with mission payload.

Figure 10: Relation between the ground distance for climb segment with mission payload.

10 of 22

American Institute of Aeronautics and Astronautics

 Figure 11: Fuel difference between Lee and Chatterji [32] assumption and BADA simulations during descent segment.

Step 5 Compute the fuel burn during the climb segment, Wclimb , by implementing the fuel increment factor fclimb :

W̃f,eqcruise + fclimb W0
Wf, climb = (5)
(1 − fclimb )
where W0 refers to final weight on climb segment. The formula above is the same as the one used by Lee and
Chatterji [32]; however, we derive the fuel factor from BADA simulation results. The fuel factor is assumed
constant for each aircraft type.
Step 6 Compute the maneuver and reserve fuel burns by using the fuel factor, following the approach by Lee and
Chatterji [32], and using their recommended values as well. The fuel increment factor is used as a fraction of
weights. The maneuver fuel weight, Wf, man , is typically expressed as a fixed fraction of the the takeoff weight
(WTO ), which can be obtained by substituting the result from Eqn. 5 to WTO = Wf, climb + W0 , with fman set to
0.007 [33]. Similarly, the reserve fuel is expressed as a fixed fraction of the zero-fuel weight (WZF ), with fres
set to 0.08 [33]. These relations are expressed below,

Wf, man = WTO · fman and Wf, res = WZF · fres (6)

Upon completing the above procedure, we obtain the fuel burn value corresponding to a flight mission. Upon
running this procedure with all flight missions in the inventory, we obtain a fuel burn database. This database provides
samples required to derive the surrogate model for each aircraft type, which is described next.
Regression modeling techniques
A regression model to approximate fuel burn is derived for each aircraft type. As mentioned in Section III, we employ
two-dimensional linear regression models, with mission payload P and range R as inputs, as fuel burn as output. This
equation is expressed as follows:
Wf,ij = αj Rij + βj Pij + γj + ij , (7)
where i denotes the flight index within an aircraft type and j denotes the aircraft type index. The variables αj , βj ,
and γj are the regression parameters for the aircraft type j, and Wf refers to the fuel burn. ij is the error term for a
specific flight, i.e., the discrepancy between the linear approximation of fuel burn W̃f,ij
and the actual fuel burn Wf,ij .
This error has a normal distribution with a mean 0 and a variance σ 2 , i.e.,  ∼ N 0, σ 2 . The method of least squares
is employed to derive αj , βj , and γj . The selected regression parameters for a specific aircraft type are denoted as aj ,
bj , and cj . Thus the approximation model can be expressed as:

W̃f,ij = aj Rij + bj Pij + cj (8)

11 of 22

American Institute of Aeronautics and Astronautics

Once the fuel burn model is derived and these parameters are obtained, we can easily find the approximated total fuel
burn for each aircraft type by simply:
Nf j Nf j
X X
Wf,j = aj Rij + bj Pij + Nf j · cj , (9)
i=1 i=1

which offers a significant computational cost reduction in estimating the total fuel burn, as the fuel burn computation
for each flight is no longer required. The prediction error of this summation will have a mean value of 0, which is
derived in the following expression:

E [Wf,j ] = E [Nf j j ] = Nf j · E [j ] = 0, (10)

since we know that E [j ] = 0.

The goodness of fit of the derived linear regression functions is evaluated by computing the coefficient of deter-
mination, or R2 . This accuracy measure indicates how well the regression line approximates the real data points. R2
values range from 0 to 1, with R2 = 1 indicates that the regression line fits the data perfectly. The derived coefficients
also quantify how much a change in each input variable would affect the total fuel burn computation.
Data Clustering
Since we want to discover groups of similar characteristics within a pool of unlabelled data by finding the data’s
hidden structure, an unsupervised learning algorithm will be used for this purpose. In particular, we look into two
algorithms, namely the classical K-means clustering [40, 41, 42], and the Gaussian mixture models (GMM) [43]. We
use the gradient information from the derived regression models, ∂Wf /∂R and ∂Wf /∂P , as inputs to the unsupervised
clustering algorithm. These quantities reflect the fuel burn performance of each aircraft type, by indicating how much,
on average, a unit change in mission range and payload affects the amount of fuel burned. By doing so, we can group
aircraft types with similar fuel burn performance together.

V. Results and Validation

In this section, we present and discuss the results of our proposed fuel burn approximation modeling procedure.
The flight mission inventories and parameter data used in the procedure are first presented, followed by the generated
fuel burn database following the procedure presented in Section IV.B. We then discuss the regression and clustering
results. The appropriate validation procedure for each method is also presented.
Flight mission inventories and parameter data
We obtain mission data corresponding to 37 distinct aircraft types from the US Department of Transportation Bureau
of Transportation Statistics (BTS) [44]. The flights included are those going to and from the United States within
the year of 2015. The aircraft types and their corresponding numbers of flight missions considered in this work are
tabulated in Table 3. In addition to the flight missions, we also need to obtain some parameters corresponding to each
aircraft and engine type to enable computing the missions’ fuel burn. At this stage, we assume the same engine type
for each aircraft type for simplicity, and the engine TSFC is assumed constant throughout the flight mission. Note that
it might not be true in real cases, and the engine variation can be included in the future to further refine our analyses. A
summary of all parameters assumed and used in this work, and their corresponding data sources, are given in Table 1.
Fuel burn database
For each aircraft type, a fuel burn database is generated by performing the fuel burn computation procedure for each
flight mission belonging to this group. The same aircraft and engine parameters are used for all flight missions of the
same aircraft type. The generated database will then be used to derive the corresponding regression model.
As mentioned, we first need to derive some information based on BADA trajectory simulation results. This infor-
mation includes:
1. The range for descent segment, Rdescent .
As mentioned in Section IV.B, we first need to identify the aircraft types that exhibit linear relationship between
payload and Rdescent , and those with constant Rdescent values. Table 2 shows the list of aircrafts belonging to each
category.
2. The fuel factor corresponding to the descent segment, fdescent , and the constant C from Equation 3.
3. The range for climb segment, Rclimb , which is expressed as a linear function of payload.

12 of 22

American Institute of Aeronautics and Astronautics

 Discipline Parameter Source
Flight mission Payload
Range Bureau of Transportation Statistics (BTS)
Flight schedule
Aircraft Maximum zero-fuel weight
Cruise Mach number
Cruise altitude
Maximum seat capacity Aircraft manufacturers (e.g., Boeing and Airbus) websites,
Climb and descent speeds http://www.airliners.net, and
The lift-to-drag ratio, L/D http://www.skybrary.aero
Takeoff distance
Landing distance
Engine TSFC Engine database handbook and websites

Table 1: List of parameters used in the fuel burn computation and their sources.

4. The fuel factor corresponding to the climb segment, fdescent , which is used in Equation 5.
These values are shown in Figure 12, where each aircraft type is indicated by a different color.

Figure 12: Fuel increment factor for the climb segment using BADA information as a function of payload.

We use the chi-square hypothesis test with a confidence level of 95% to validate the sufficiency of the number of
samples. We randomly select 10 samples for mission payload including the lower and upper bounds of the BADA
information for each aircraft type.
The chi-square hypothesis is validated with the coefficient of variation (cv ), to validate the variability in relation
to the mean of the population. This coefficient can be defined as ratio of standard deviation σ (or s if using sample
standard deviation) to the mean µ (or x̄ if using sample mean) [45, 46]. However, we use an approach to use an
unbiased estimator which is more suitable for our small sample size [47]. The unbiased estimator can be expressed as:
 
1
cˆv ∗ = 1 + cˆv (11)
4n

All aircraft types have cˆv ∗ below 3%, therefore we can assume that the data are good enough and considered as
consistent [48]. Huber [49] defined an estimator between mean and median which is claimed to be more robust. This

13 of 22

American Institute of Aeronautics and Astronautics

No. Aircraft with constant range Aircraft with linear function range
1 Airbus Industrie A319 Canadair RJ-100/100ER
2 Airbus Industrie A320-100/200 Canadair RJ-200ER/RJ-440
3 Airbus Industrie A330-200 Canadair RJ-700
4 Airbus A330-300 Canadair RJ-900
5 Airbus Industrie A318 Embraer 140
6 Airbus Industrie A321 Embraer 145
7 Airbus Industrie A340-200 Embraer 170
8 Airbus Industrie A340-300
9 Airbus Industrie A340-500
10 Airbus Industrie A340-600
11 Boeing 737-100/200
12 Boeing 737-300
13 Boeing 737-400
14 Boeing 737-500
15 Boeing 737-600
16 Boeing 737-700/700LR
17 Boeing 737-800
18 Boeing 737-900
19 Boeing 737-900ER
20 Boeing 757-200
21 Boeing 757-300
22 Boeing 777-200ER/200LR/233LR
23 Boeing 777-300/300ER/333ER
24 Boeing 747-200/300
25 Boeing 747-400
26 Boeing 767-200/ER/EM
27 Boeing 767-300/300ER
28 McDonnell Douglas DC9 Super 80/MD81/82/83/88
29 McDonnell Dougles MD-90
30 Embraer 190
Table 2: List of aircraft in constant range category and linear function range category for descent segment.

14 of 22

American Institute of Aeronautics and Astronautics

estimator is called the M-estimator or the Huber-estimator [46]. However, while comparing the mean, median, and
huber-estimator values, those values show negligible difference between each other. Therefore, the constant value of
fclimb is estimated with mean value since the mean value is more commonly used.
In addition to deriving the aforementioned parameters based on BADA data, we also need to derive the ratio of
climb-descent distances (RCD) for short-haul flights. The RCD values are derived by implementing the triangle-
similarity theorem. Figure 13 plots the RCD as a function of payload weight for each aircraft type (color-coded). Only
long-haul flights are included in these plots, where each dot represents a long-haul flight mission. Figure 13 suggests
a linear relationship between RCD and payload. Upon deriving this relationship for each aircraft type, we can then
calculate the climb and descent ranges for short-haul missions.

Figure 13: Climb-descent ratio is proportional with payload.

Figure 14 shows the distribution of fuel burn for different mission range (in the x-axis) and payload (indicated by
the colormap). Only three representative aircraft types (the Boeing 777-200ER, Airbus 320-100, and the McDonnell
Douglas DC9 Super80) are shown here, considering the space constraint. However, similar trends are observed on all
other 34 remaining aircraft types, though with different “sensitivity” of fuel burn with respect to payload and range,
as will be further discussed next.

(a) Boeing 777-200ER (b) Airbus 320-100 (c) McDonnell Douglas DC9

Figure 14: Fuel burn plots as functions of mission range and payload for three sample aircraft types.

Regression models
The generated fuel burn data shown in Figure 14 suggest linear relationships with mission payload and range. For
each aircraft type, a regression model is derived to approximate each mission’s fuel burn as a function of these two
mission specification parameters. Table 3 lists the derived linear regression equation. The corresponding R2 values

15 of 22

American Institute of Aeronautics and Astronautics

for each aircraft is also displayed to indicate the goodness of fit of each regression function. We can observe that all
R2 values are greater than 0.98, which tells us that the linear relationship assumption is valid, and that the total fuel
burn can be sufficiently explained by the selected two input variables. The hundreds or thousands of flight missions
considered for each aircraft type has distinct mission specifications, and yet the generated fuel burn data show that
there is a “pattern” within each aircraft type. From the regression coefficients shown in Table 3, we can see that the
flight’s fuel burn depends largely on the mission range, and much less so on the mission payload.
No. Aircraft Name No. of Flights Regression Model R2
1 Airbus Industrie A319 24 113 Wf = 1.809R + 0.145P + 580.83 0.995
2 Airbus Industrie A320-100/200 29 175 Wf = 1.984R + 0.138P + 390.97 0.997
3 Airbus Industrie 330-200 4 005 Wf = 8.305R + 0.341P − 2314.39 0.996
4 Airbus A330-300 1 327 Wf = 8.278R + 0.329P + 706.80 0.997
5 Boeing 737-100/200 341 Wf = 2.085R + 0.148P + 125.44 0.996
6 Boeing 737-300 11 243 Wf = 2.099R + 0.146P + 470.24 0.996
7 Boeing 737-400 5 403 Wf = 2.412R + 0.16P + 431.29 0.99
8 Boeing 737-500 790 Wf = 1.825R + 0.117P + 631.40 0.996
9 Boeing 737-600 224 Wf = 1.922R + 0.185P + 517.48 0.998
10 Boeing 737-700/700LR 26 228 Wf = 2.063R + 0.165P + 915.72 0.996
11 Boeing 737-800 33 809 Wf = 2.539R + 0.157P + 484.68 0.996
12 Boeing 737-900 10 171 Wf = 2.507R + 0.172P + 503.32 0.997
13 Boeing 737-900ER 821 Wf = 2.547R + 0.18P + 328.96 0.997
14 Boeing 757-200 10 256 Wf = 2.966R + 0.168P + 443.28 0.997
15 Boeing 757-300 1 857 Wf = 3.242R + 0.177P + 762.16 0.997
16 Boeing 777-200ER/200LR/233LR 4 732 Wf = 6.724R + 0.396P − 7634.59 0.992
17 Boeing 777-300/300ER/333ER 1 791 Wf = 8.163R + 0.371P − 8576.72 0.994
18 Canadair RJ-200ER /RJ-440 20 641 Wf = 1.045R + 0.116P + 140.42 0.98
19 Canadair CRJ 900 14 586 Wf = 1.271R + 0.141P + 778.27 0.992
20 Embraer-145 17 974 Wf = 0.705R + 0.134P + 601.88 0.995
21 McDonnell Douglas DC9 Super 13 976 Wf = 2.463R + 0.152P − 67.81 0.993
80/MD81/82/83/88
22 Airbus Industrie A-318 123 Wf = 1.407R + 0.172P + 17.29 0.999
23 Airbus Industrie A321 4 767 Wf = 2.61R + 0.176P + 880.53 0.997
24 Airbus Industrie A340-300 372 Wf = 8.81R + 0.424P − 5057.09 0.997
25 Airbus Industrie A340-500 146 Wf = 10.86R + 0.453P − 9708.52 0.996
26 Airbus Industrie A340-200 506 Wf = 8.36R + 0.348P − 290.49 0.996
27 Airbus Industrie A340-600 465 Wf = 12.428R + 0.385P − 8552.62 0.995
28 Boeing 747-200/300 129 Wf = 10.06R + 0.403P − 8964.48 0.998
29 Boeing 747-400 1 789 Wf = 8.779R + 0.247P − 9191.53 0.994
30 Boeing 767-200/ER/EM 315 Wf = 4.102R + 0.228P + 716.20 0.995
31 Boeing 767-300/300/300ER 6455 Wf = 5.159R + 0.279P − 2282.61 0.993
32 Canadair RJ-700 18 455 Wf = 1.194R + 0.143P + 725.06 0.993
33 Canadair RJ-100/RJ-100ER 185 Wf = 0.781R + 0.169P + 351.56 0.998
34 Embraer-140 2645 Wf = 0.673R + 0.133P + 633.03 0.997
35 Embraer 170 7 644 Wf = 1.19R + 0.144P + 949.164 0.998
36 Embraer 190 5 756 Wf = 1.626R + 0.145P + 1219.34 0.996
37 McDonnell Douglas MD-90 2 619 Wf = 2.157R + 0.117P + 643.18 0.996
Table 3: The list of aircraft type, the number of flight missions included to generate the fuel burn database, and the derived two-dimensional linear regression models
for each aircraft type.

Aircraft type clustering

While all 37 aircraft types show that their fuel burns vary linearly with payload and range, how much they vary is
different for each aircraft type. These input sensitivity is quantified in the regression gradient, i.e., ∂Wf /∂R (with
respect to range) and ∂Wf /∂P (with respect to payload). We can then use this sensitivity information to study the
fuel burn performance patterns of the different aircraft types, and use this knowledge to group similar aircraft types
together. The clustering is performed using two unsupervised algorithms, i.e., GMM and K-means methods. Two

16 of 22

American Institute of Aeronautics and Astronautics

clusters were considered with GMM clustering, and three clusters were attempted for both algorithms. The clustering
results show that the algorithm can automatically group aircraft types with similar sizes together. The three clusters
from the K-means clustering, for instance, separate aircraft types with seat capacity more than 300 (e.g., Airbus A330,
Airbus A340, Boeing B747, Boeing B777), those with seat capacity between 200 and 300 (e.g., Boeing 757, Boeing
767), and those with seat capacity less than 200 (e.g., Airbus A320, Boeing 737, Embraer and Canadair regional
aircraft). The automatic clustering algorithm would be even more useful when there are a large number of aircraft
types included in the analysis, where manual classification would be challenging.

(a) GMM with 2 clusters (b) GMM with 3 clusters (c) K-means with 3 clusters

Figure 15: Clustering results using the Gaussian mixture models (GMM) and K-means clustering algorithms, using the gradients of fuel burn with respect to distance
and payload as the clustering criterion.

Model validation
In this section, we first look into the fuel burn estimation errors for the short-haul approach when the cruise-only
assumption is adopted. The computed fuel burn values will be compared to those obtained from assuming a saw-tooth
profile.
Figure 16 shows three representative aircraft types (Airbus A321, Boeing B737-900, and Canadair RJ700) in
the fuel burn distribution for short-haul mission. The blue dots represent the data corresponding to the cruise-only
assumption, whereas the red triangles correspond to the saw-tooth profile assumption. This result clearly shows the
large discrepancies between the two approaches, which suggests that we could incur a significant approximation error
when the wrong assumption is adopted.

(a) Airbus A321 (b) Boeing B737-900 (c) Canadair RJ700

Figure 16: Fuel consumption difference between the cruise-only assumption and saw-tooth profile for short-haul missions.

To quantify the observed discrepancies, we compute the corresponding percentage RMSD errors. Figure 17 shows
the RMSD for all aircraft types using the available BTS mission data from the years of 2015 (blue) and 2016 (red).
Note that some aircraft types do not have any relevant data for the year of 2016, and thus they are omitted in the
plot. The minimum discrepancy is 22.07%, and some aircraft types have more than 100% errors. The cruise-only
approximation, which does not reflect the actual aircraft operation and yet is often adopted in fuel burn estimation,
can therefore lead to large errors, which can propagate to any subsequent analyses. This is undesirable and further
emphasizes the importance of making the more realistic assumption in the modeling process.

17 of 22

American Institute of Aeronautics and Astronautics

 Figure 17: RMSD between the saw-tooth and cruise-only mission profiles in short-haul missions for all aircraft types.

The derived regression fuel burn model is validated using data obtained from the BTS within the period of January
to October 2016. The error between the fuel burn computed using the segment-by-segment fuel burn computation
procedure and the derived regression model is quantified using the root mean square deviation (RMSD). We perform
two types of validations to measure the error. The first is to validate the accuracy of the prediction for each flight
mission data, i.e., for each data point. This validation is performed using 10 sets of 100 randomly selected samples
and will be denoted as the “individual sampling” validation. The percentage error is then calculated using the average
RMSD from all repetitions, and the standard error is computed from the standard deviation corresponding to each
repetition. Secondly, we calculate the percentage RMSD of the aggregate fuel burn computation, denoted as the
“aggregate” validation. For this purpose, we take the summation of all computed fuel burn corresponding to each
aircraft type from BTS within period of January to October 2016. We then obtain the difference between the total
aggregate fuel burn computed after performing the fuel burn for each flight mission one-by-one, and by applying
Equation 9, i.e., the regression model. Figure 18 shows the percentage error (RMSD) for both conditions and the
standard error for individual sampling condition for all aircraft types considered in this study.
We observe from Figure 18 that the percentage error (RMSD) for the individual sampling validation is at least
one order of magnitude higher than that of the aggregate validation. The maximum values for them are around 10%
and 0.65%, respectively. This result is not surprising, and is consistent with the predicted expected aggregate error as
shown in Equation 10. From this result, we can confidently use the derived regression model to accurately estimate
the total fuel burn corresponding to an aircraft type, thereby significantly reducing the computational cost.

VI. Conclusion
In this paper, we have shown that we can derive a two-dimensional linear regression model to express the fuel burn
of each aircraft type as a function of mission range and payload. The accuracy of such a simple model was very high,
with R2 more than 98% for all aircraft types, which suggested that the linear assumption was valid, and that the two
input variables were sufficient to describe the fuel burn performance. The derived model is particularly useful when
we need to approximate the total aggregate fuel burn, given a list of flight missions. Using the regression models,
computing the total aggregate fuel burn becomes a function of the number of aircraft types,

instead of the number of
flight missions. This could provide a computational cost reduction in the order of O 104 , which would be significant.
The validation procedure, which was done with over 30 aircraft types, showed that the surrogate model could predict
the total aggregate fuel burn with less than 1% approximation error. These results showed that the models were both
efficient and accurate, and thus we could use them confidently in any subsequent decision-making and policy analyses.
For instance, the proposed models could provide fast and yet accurate approximations of the environmental impacts
and fuel consumptions of the different policy scenarios, which could help policy makers making their decision.
The derivation of the fuel burn regression models required the fuel burn database corresponding to each aircraft
type, since we used a sample-based surrogate modeling technique. For this purpose, we derived a medium-fidelity,
data-enhanced method to evaluate the fuel burn for each flight mission. We consulted the US Department of Trans-
portation’s Bureau of Transportation Statistics (BTS) for the actual flight mission data to generate the fuel burn

18 of 22

American Institute of Aeronautics and Astronautics

(a) RMSD for individual sampling (b) Standard error for individual sampling

(c) RMSD for aggregate

Figure 18: Percentage error between regression fuel burn model and fuel burn computation.

19 of 22

American Institute of Aeronautics and Astronautics

database. By using actual data, the derived regression models would reflect the actual fuel burn performance of
each aircraft type in real world applications.
We first needed to obtain the corresponding fuel burn for each of the flight mission included in this database.
For this purpose, we developed a medium-fidelity, data-enhanced fuel burn computation procedure. We modeled
the mission profile to include the takeoff, climb, cruise, descent, and landing for long-haul flights. For short-haul
flights, the mission profile only inlcuded takeoff, climb, descent, and landing; the aircraft would never reach its cruise
altitude. This mission profile resembled how aircraft actually operated more closely, and thus the results would be more
accurate than if we simplified the fuel burn computation by assuming only cruise segment. The fuel burn computation
comparison showed that assuming a cruise-only mission profile grossly overestimates the actual fuel burn for short-
haul flights, with discrepancies of more than 100%. This would lead to undesirable results should the values were to be
used in any subsequent analyses or as a basis for decision making. An expression was derived for each mission segment
(climb, cruise, and descent), which was based on the classical Breguet range equation. This equation summarizes the
aerodynamic discipline (through the L/D), structures (through the weights), and propulsion (through the TSFC), and
is commonly used to evaluate aircraft performance, including the fuel burn. A correction factor was derived for each
of the climb and descent segment, since the Breguet range equation contains some assumptions that deem it invalid
for these two segments. The correction factors were derived for each aircraft type based on the BADA trajectory
simulation results.
In addition to the fuel correction factors, we also used the BADA trajectory simulation results to obtain the climb
and descent distances. While the climb distance tended to be constant for each aircraft type, we found for smaller
aircraft, the descent distance exhibited a linear relationship with the mission payload, while for others the descent
distance was mainly constant. The aircraft categorization was done by employing the K-means clustering algorithm.
For each aircraft type, we used 10 BADA samples, which were deemed sufficient by the chi-square hypothesis test.
The fuel burn regression model derived for each aircraft type provided information on the sensitivity of the fuel
burn performance with respect to mission payload and range, i.e., the regression gradients. This information could then
be used to identify aircraft types that have similar fuel burn performance, and group them together. This classification
was done automatically by employing an unsupervised learning algorithm. In this work, we demonstrated the use of
the Gaussian mixture model (GMM) and K-means clustering to produce 2 and 3 clusters of aircraft types. We could
see that in general, similar-size aircraft types were grouped together.
For the future work, we can include more aircraft types in the analyses to represent the actual aircraft-type distri-
bution in the current air traffic. Moreover, we can also consider the different engine types that are used for the same
aircraft type. The different engines might result in different fuel burn performance, which could be reflected by a
higher-dimensional linear regression model (e.g., by including the engine TSFC as one of the input variable), or by
using a more complex surrogate model. With more units (i.e., more aircraft types or more aircraft-engine combina-
tions), the automatic clustering algorithm would be even more useful in identifying similar units together, as it would
become more challenging to do it manually. Once the clusters are identified, we can derive a “higher-level” fuel burn
surrogate model, one for each cluster, by first identifying the appropriate surrogate modeling technique to be used.
When we use these surrogate models in the total aggregate fuel burn approximation, this computation would now be
a function of the number of clusters, which is significantly smaller than the number of flight missions to be evaluated.
This would provide the desired computational efficiency in decision-making and policy analyses.

References
[1] Waitz, I., Lukachko, S., Go, Y., Hollingsworth, P., Harback, K., and Morser, F., “Requirements Document for the Aviation
Environmental Portfolio Management Tool (APMT),” Partnership for Air Transportation Noise and Emissions Reduction
(PARTNER), Report No.: PARTNER-COE-2006-001, 2006.
[2] ICAO, “Aviation and Climate Change,” International Civil Aviation Organization (ICAO) Environmental Report, 2010.
[3] International Energy Agency (IEA), “World Energy Outlook 2008,” International Energy Agency, Paris, France, 2008.
[4] ICAO, “On board a sustainable future,” International Civil Aviation Organization (ICAO) Enviromental Report, 2016.
[5] Lee, J. J., Historical and Future Trends in Aircraft Performance, Cost, and Emissions, Master’s thesis, Aeronautics & Astro-
nautics Department and Technology & Policy Program, Massachusetts Institute of Technology, September 2000.
[6] Brunelle-Yeung, E., The impacts of aviation emissions on human health through changes in air quality and UV irradiance,
Master’s thesis, Massachussets Institute of Technology, 2009.
[7] for Computer Applications in AeroSpace Science, C. and Engineering, “Top Level Technology in Numerical Simulation for
Aircraft Design,” PowerPoint presentation, Deutsches Zentrum für Luft- und Raumfahrt (DLR), 2010.
[8] Wynne, R., “Geographic Dimensions of Airline Network Development,” Presentation at the Association of American Geog-
raphers (AAG) in Seattle, 2011.
[9] Bloomberg, “Why the Longest Nonstop Flight are Ending,” http://www.bloomberg.com/news/articles/
2013-10-31/fuel-costs-lead-singapore-air-to-end-worlds-longest-nonstop-flights (last
accessed: October 17, 2016).

20 of 22

American Institute of Aeronautics and Astronautics

[10] Liem, R. P., Mader, C. A., Lee, E., and Martins, J. R. R. A., “Aerostructural design optimization of a 100-passenger regional
jet with surrogate-based mission analysis,” AIAA Aviation Technology, Integration, and Operations (ATIO) Conference, Los
Angeles, CA, August 2013. doi:10.2514/6.2013-4372.
[11] Liem, R. P., Mader, C. A., and Martins, J. R. R. A., “Surrogate Models and Mixtures of Experts in Aerody-
namic Performance Prediction for Mission Analysis,” Aerospace Science and Technology, Vol. 43, 2015, pp. 126–151.
doi:10.1016/j.ast.2015.02.019.
[12] Randle, W. E., Hall, C. A., and Vera-Morales, M., “Improved Range Equation Based on Aircraft Flight Data,” Journal of
Aircraft, Vol. 48, No. 4, July–August 2011, pp. 1291–1298. doi:10.2514/1.C031262.
[13] Kenway, G. K. W., Kennedy, G. J., and Martins, J. R. R. A., “A Scalable Parallel Approach for High-Fidelity Aerostructural
Analysis and Optimization,” 53rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference,
Honolulu, HI, April 2012, AIAA 2012-1922.
[14] Liem, R. P., Kenway, G. K. W., and Martins, J. R. R. A., “Multimission Aircraft Fuel Burn Minimization via Multipoint
Aerostructural Optimization,” AIAA Journal, Vol. 53, No. 1, 2015, pp. 104–122. doi:10.2514/1.J052940.
[15] Simos, D. and Jenkinson, L. R., “Optimization of the Conceptual Design and Mission Profiles of Short-Haul Aircraft,” Journal
of Aircraft, Vol. 25, No. 7, July 1988, pp. 618–624.
[16] Diedrich, A., Hileman, J., Tan, D., Willcox, K., and Spakovsky, Z., “Multidisciplinary Design and Optimization of the Silent
Aircraft,” 44th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, January 9–12 2006. doi:10.2514/6.2006-1323,
AIAA 2006-1323.
[17] Waitz, I., Lukachko, S., Willcox, K., Belobaba, P., Garcia, E., Hollingsworth, P., Mavris, D., Harback, K., Morser, F., and
Steinbach, M., “Architecture Study for the Aviation Environmental Portfolio Management Tool,” Partnership for Air Trans-
portation Noise and Emissions Reduction PARTNER, Report No.: PARTNER-COE-2006-002, 2006.
[18] Liem, R. P., Multimission Fuel-Burn Minimization in Aircraft Design: A Surrogate-Modeling Approach, Ph.D. thesis, Univer-
sity of Toronto, 2015.
[19] Yan, B., Jansen, P. W., and Perez, R. E., “Multidisciplinary Design Optimization of Airframe and Trajectory Considering
Cost and Emissions,” 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization (MAO) Conference, Indianapolis, IN,
September 2012. doi:10.2514/6.2012-5494, AIAA 2012-5494.
[20] Lee, J. J., Lukachko, S. P., Waitz, I. A., and Schafer, A., “Historical Future trends in aircraft performance, cost, and emissions,”
Annual Review of Energy and the Environment, Vol. 26, November 2001, pp. 167–200. doi:10.1146/annurev.energy.26.1.167.
[21] Graham, W. R., Hall, C. A., and Vera-Morales, M., “The potential of future aircraft technology for noise and pollutant
emissions reduction,” Transport policy, Vol. 34, 2014, pp. 36–51. doi:10.1016/j.tranpol.2014.02.017.
[22] Singh, V. and Sharma, S. K., “Fuel consumption optimization in air transport: a review, classification, critique, sim-
ple meta-analysis, and future research implications,” European Transport Research Review, Vol. 7, No. 12, 2015, pp. 24.
doi:10.1007/s12544-015-0160-x.
[23] Coffin, J. G., “A Study of Airplane Range and Useful Loads,” NACA-TR-69, NACA, 1920.
[24] Breguet, L., “Calcul du Poids de Combustible Consummé par un Avion en Vol Ascendant,” Comptes Rendus Hebdomodaires
des Séances de l’Académie des Sciences, Vol. 177, July 1923, pp. 870–872.
[25] McCormick, B. W., Aerodynamics, Aeronautics, and Flight Mechanics, John Wiley & Sons, New York, US, 1979.
[26] Dancila, B. D., Botez, R., and Labour, D., “Fuel burn prediction algorithm for cruise, constant speed and level flight segments,”
Aeronautical Journal, Vol. 117, No. 1191, 2013, pp. 491–504. doi:10.1017/S0001924000008149.
[27] Roskam, J., Airplane Design Part I: Preliminary Sizing of Airplanes, Roskam Aviation and Engineering Corporations, Ottawa,
KS, 1985.
[28] O’Kelly, M. E., “Fuel burn and environmental implications of airline hub networks,” Transportation Research Part D: Trans-
port and Environment, Vol. 17, No. 7, 2012, pp. 555–567.
[29] Nuic, A., Poles, D., and Mouillet, V., “BADA: An advanced aircraft performance model for present and future ATM systems,”
International Journal of Adaptive Control and Signal Processing, Vol. 24, 2010, pp. 850–866. doi:10.1002/acs.1176.
[30] Base of Aircraft Data (BADA) Aircraft Performance Modelling Report: EEC Technical Report No. 2009-009, April 2009.
[31] Dalmau, R. and Prats, X., “Fuel and time savings by flying continuous cruise climbs Estimating the benefit pools for maximum
range operations,” Transportation Research Part D, Vol. 35, 2015, pp. 62–71. doi:10.1016/j.trd.2014.11.019.
[32] Lee, H. and Chatterji, G. B., “Closed-Form Takeoff Weight Estimation Model for Air Transportation Simulation,” 10th AIAA
Aviation Technology, Integration, and Operations (ATIO) Conference, Fort Worth, TX, Sept 13–15 2010. doi:10.2514/6.2010-
9156, AIAA 2010-9156.
[33] Kroo, I. M., Aircraft Design: Synthesis and Analysis, Desktop Aeronautics, Palo Alto, CA, 1st ed., Sept 2006.
[34] Giannakoglou, K. C., Papadimitriou, D. I., and Kampolis, I. C., “Aerodynamic shape design using evolutionary algorithms and
new gradient-assisted metamodels,” Computer Methods in Applied Mechanics and Engineering, Vol. 195, 2006, pp. 6312–
6329. doi:10.1016/j.cma.2005.12.008.
[35] Simpson, T. W., Booker, A. J., Ghosh, D., Giunta, A. A., Koch, P. N., and Yang, R. J., “Approximation methods in multi-
disciplinary analysis and optimization: A panel discussion,” Structural and Multidisciplinary Optimization, Vol. 27, 2004,
pp. 302–313.
[36] Sobieszczanski-Sobieski, J. and Haftka, R. T., “Multidisciplinary aerospace design optimization: Survey of recent develop-
ments,” Structural Optimization, Vol. 14, 1997, pp. 1–23. doi:10.1007/BF01197554.
[37] Stigler, S. M., “Gergonne’s 1815 paper on the design and analysis of polynomial regression experiments,” Historia Mathe-
matica, Vol. 1, No. 4, 1974, pp. 431–439. doi:10.1016/0315-0860(74)90034-2.
[38] NASA, “Information Summaries Sawtooth Climbs,” IS-97/08-DFRC-SAW1 http://services.aero.polimi.it/
˜chimetto/bacheca/downloads/docs/NASA03SawtoothClimb.pdf.

21 of 22

American Institute of Aeronautics and Astronautics

 [39] Lee, J. J., “Can we accelerate the improvement of energy efficiency in aircraft systems?” Energy Conversion and Management,
Vol. 51, 2010, pp. 189–196.
[40] Steinhaus, H., “Sur la division des corps materiels en parties,” Bull. Acad. Polon. Sci., Vol. 1, 1956, pp. 801–804.
[41] Lloyd, S. P., “Least Squares Quantization in PCM,” IEEE Transactions on Information Theory, Vol. 28, No. 2, March 1982,
pp. 129–137. doi:10.1109/TIT.1982.1056489.
[42] MacQueen, J., “Some methods for classification and analysis of multivariate observations,” Fifth Berkeley Symposium on
Mathematics. Statistics and Probability., University of California Press., 1967, pp. 281–297.
[43] Reynolds, D., “Gaussian Mixture Models,” Encyclopedia of Biometric Recognition, Springer, February 2008.
[44] U.S. Department of Transportation, R. and Innovative Technology Administration, B. o. T. S., National Transportation Statis-
tics: http://www.bts.gov/publications/national_transportation_statistics/.
[45] Montgomery, D. C. and Runger, G. C., Applied Statistics and Probability for Engineers, John Wiley & Sons, Inc., 2003.
[46] Casella, G. and Berger, R. L., Statistical Inference, Duxbury, 2nd ed., 2001.
[47] Sokal, R. R. and Rohlf, F. J., Biometry, Freeman, 4th ed., 2012.
[48] Zady, M. F., “Z-4: Mean, Standard Deviation, and Coefficient of Variation,” https://www.westgard.com/
lesson34.htm [last accessed: April 22, 2017].
[49] Huber, P. J., “Robust Estimation of Location Parameter,” The Annals of Mathematical Statistics, Vol. 35, No. 1, 1964, pp. 73–
101.

22 of 22

American Institute of Aeronautics and Astronautics

View publication stats