Flight plans, but greener: The ICCT and Google’s mission to refine the Travel Impact Model

In seeking to advance sustainability in the aviation industry, robust estimates of emissions are essential because they support informed decision-making. Previous ICCT research found that travelers on U.S. routes can reduce carbon dioxide (CO₂) emissions attributable to their ticket by 22% on average, and up to 63%, by choosing the least-emitting flight available.  

Consumers can access emissions estimates for future flights around the world from Google’s Travel Impact Model (TIM), and since 2023, the ICCT has worked with global experts to refine the model through an Advisory Committee that we lead. This partnership established seven core principles and a roadmap to ensure continuous improvement of the TIM for rigorous, transparent, and consistent emissions reporting.  

As the TIM continues to evolve, it holds promise to become the global standard in low-carbon travel search. Here I’ll highlight some of the work behind its development, including the model validation and model selection approaches that enhance its reliability for travelers, airlines, and policymakers.  

Estimating fuel burn is the first step in assessing emissions, and this is challenging due to the variability in flight operations. Emissions will differ based on aircraft technology, weather, and operational practices, and even for the same aircraft and route, fuel burn can vary significantly. The TIM uses the European Environment Agency (EEA) model, which bases estimates on aircraft type and route distance and creates simplified linear relationships. Using the operations of Brazilian airlines in 2019 and the four most commonly used aircraft types, Figure 1 compares the TIM (version 1.8.0) estimates (dashed lines) with real-world fuel burn at the flight level (dots). The large variability reflects the complexity and uncertainty of flight operations, where fuel burn depends on a range of interdependent and sometimes unpredictable factors such as weather and operational practices.
 

Figure 1. Fuel burn versus distance for each individual flight by the four most common aircraft types in the ANAC data in 2019

Model validation

Validation is a quantitative assessment of how well model prediction represents real-world data. For validating the TIM, we needed data from past flights, including fuel burn, ideally at the flight level or at least aggregated by route and aircraft type. The only public dataset identified that met the requirements was the Brazilian Civil Aviation Agency (ANAC) microdata, which provides historical flight data for Brazil since the year 2000 at the flight level. As it’s limited to Brazilian airlines, Google combined ANAC’s public data with private operational data shared by partner airlines worldwide. The aircraft types covered by this sample represent approximately 76% of global flights in 2019 and the validation sample now contains more than 3 million flights. The Google engineering team is continuously working with airlines to expand it to enhance model representativeness and reliability. There is a three-step process to promote reliability: 

1. Data cleaning: We remove irrelevant or incomplete data. 
2. Data aggregation: We group fuel burn data by route, aircraft type, and airline. This is necessary because some private airline data was shared in aggregated form; it contained fuel burn averages by route and aircraft type rather than at the flight level. By aligning our analysis with the level of granularity available in the shared data, we ensure consistency. 
3. Error analysis: We compare the TIM’s estimates with real-world fuel burn using metrics such as median absolute error and error distribution. “Error” is defined as the difference between actual and estimated fuel burn, expressed as a percentage. The actual fuel burn refers to the values in the validation dataset, and estimated fuel burn refers to the TIM estimates. Positive errors indicate overestimation and negative errors indicate underestimation. 

The TIM validation framework uses four key metrics for evaluation: median absolute error (the central tendency of errors), error threshold analysis (the percentage of estimates within different error bounds), distance-based metrics (error trends by route length), and the distance and aircraft error metric (error trends by route length and aircraft type). Details of the metrics are in this technical brief, and Figure 2 illustrates the error distribution curve for the TIM (version 1.8.0) estimates. As shown, the fuel burn is more often underestimated than overestimated by the model. The TIM underestimates the fuel burn for nearly 75% of the airline-aircraft type-route combinations in the validation sample.

Figure 2. Frequency (left) and cumulative (right) distributions of the error in the TIM’s fuel burn estimates when compared with the real-world fuel burn from the combination of ANAC 2019 and private airlines data

Model selection 

The TIM fuel burn estimation was originally based on the EEA 2019 model, which allows users to define only aircraft type and stage length; other significant factors like flight trajectory and payload are not included. Recognizing these limitations, the TIM Advisory Committee explored alternative fuel burn models. 

Nine models were assessed qualitatively (details in the technical brief), and five were shortlisted for detailed evaluation using the validation methodology: EEA 2023, OpenAP, Poll-Schumann, Piano 5, and ICAO ICEC. Because these models vary in structure and require different operational assumptions such as trajectory and payload, we standardized assumptions where possible to be able to compare them. The tested scenarios, based on real-world operations and described in the technical memo, reflect these simplifications. Figure 3 illustrates how the error distributions of these models compare with EEA 2019.  

Both EEA 2019 and EEA 2023 showed narrow error distributions, reflecting good accuracy. However, EEA 2023 consistently outperformed EEA 2019 across key metrics. In contrast, OpenAP demonstrated a wider error spread, indicating lower predictive accuracy for the data used. Intermediate performers, such as ICAO ICEC, Poll-Schumann, and Piano 5, showed moderate error variability. These evaluations showed EEA 2023 to be the most suitable model, and it was adopted in mid-2024. 

Figure 3. Comparison of the error distribution across alternative models

In January 2024, the Advisory Committee incorporated validation into the TIM workflow to evaluate fuel burn model updates. Then, in June 2024, in addition to adopting EEA 2023, they applied a distance correction factor that enhanced the TIM’s accuracy and alignment with real-world operations. The distance correction factor refines stage length inputs by replacing Great Circle Distance with an average route distance based on real-world flight paths. This adjustment reduced the median absolute error from 7.80% to 6.30%. Future Advisory Committee work on second-order fuel burn effects like payload, engine variants, and aircraft age is expected to further improve the accuracy of the TIM and thus further improve its value for a wide range of stakeholders, including the flying public.

Special thanks to Ana Beatriz Reboucas and Jayant Mukhopadhaya for their significant contributions to the research on the TIM website.