(535c) Development of Property-Based Models for the Estimation of the Net Heat of Combustion of Aviation Fuels

The rapid development in sustainable aviation fuels (SAF) and the constantly expanding landscape of approved or in-research aviation fuel production pathways, as well as their blends, comes with several operational and safety challenges that limit the potential environmental and economic benefits of such technologies. Many of these challenges fall within the scope of quantification of the composition and properties of these fuels to ensure their accordance with specification [ASTM D1655 & D7566]. While precision methods for measuring these quantities are very well established, the associated cost and operational overhead, as well as required expertise, limit these methods' application to laboratory conditions (e.g., net heat combustion [ASTM D4809], freezing point [ASTM D2386], aromatic content [ASTM D1319 and D5769], etc.), thus constituting difficult-to-measure quantities from a real-time field application point of view. While machine learning (ML)-models based on the detailed composition of fuels can help reduce a significant amount of effort and complexity associated with measurement (e.g., for prescreening SAF candidates [Heyne (2021)]), similar limitations restrict the field application of these models, e.g., due to their dependence on complex gas chromatography techniques [Hall (2022)]. Therefore, to address this challenge, empirical or chemometric models based on property correlations are studied, which have the potential to simplify and improve the analysis of complex aviation fuels with various compositions.

This study focuses on comparing, developing, and simplifying property-based empirical or chemometric models for estimating both conventional and sustainable aviation fuels’ properties, specifically for the scope of this study focusing on models used to estimate the net heat of combustion (NHC). Such models are developed based on long-standing chemical expertise and a large data set of fuel property precision measurements. They are primarily valid for well-defined conventional jet fuels based on crude oil and that are well represented in the training data set of such models. Therefore, the generalization of these models to alternative fuels is bounded by two challenges. Firstly, considering the statistical nature of these models and new variations in the composition of SAF, such models could become stale and, unless (re)-trained to the new variation, lose accuracy [Vozka (2018), Edwards (2023)]. Secondly, these models require the precision measurement of difficult-to-measure properties such as distillation curves or aromatic content, which make their field application difficult (Example Models: net heat of combustion [ASTM D3338], flash point [Alqaheem and Riazi (2017)], hydrogen content [ASTM D3343], calculated cetane index [ASTM D976 and D4737], etc.).

Within this scope, this study will focus on property-based models for NHC estimation that use the correlations of NHC to different aviation fuel properties. NHC is chosen due to its potential to motivate both environmental and economic incentives, which can assert the urgency of SAF development. Furthermore, due to the quantity’s correlation to both (bulk) properties and the fuel’s composition, NHC lends itself as an excellent example to illustrate the approaches developed in this study for different fuel types. As a starting point to improve and possibly simplify the analysis of NHC, the standard test method for estimating NHC, as given by ASTM D3338, is used as a reference model. This model estimates NHC based on empirical equations using precision measurement of density at 15°C, aromatic content, distillation curve (average of 10%, 50%, and 90% volume points), and sulfur content. Consequently, different alternative ML-based modeling approaches to the empirical equations of the ASTM are assessed. Finally, to address application challenges regarding the difficult-to-measure quantities required by the ASTM model, a number of simplified ML models that rely only on the bulk property as input parameters are studied. In summary, as shown in the attached model comparison diagram, the study evaluates three modeling approaches for estimating the NHC:

• The ASTM D3338 based empirical model
• The ML-based adaptions of ASTM D3338
• The simplified bulk property-based ML-models

The evaluation of the empirical model uses the extensive database of the German Aerospace Center (DLR) to evaluate and compare ASTM D3338 NHC estimations to precision methods [ASTM D4809]. The database has been extensively cited in several publications [e.g., for detailed composition models Hall (2021 & 2022)]. It encompasses precision composition and property measurements of jet fuels ranging from conventional fuel (Jet A, Jet A-1, JP-8, etc.) to different alternative fuels (HEFA, ATJ, FT-SPK, etc.) and their blends (This study uses a subset of the database, including over 5000 jet fuels, to train and evaluate all modeling approaches). Subsequently, the evaluation of the empirical approach evaluates NHC estimations of different fuel types and the sensitivity of the ASTM D3338 model to the different input parameters as well as fuel composition variations. Based on these results, the approach establishes a baseline for data-set-related uncertainty of the model that can be used to evaluate the other modeling approaches.

The evaluation of ML-based adaption of the ASTM D3338 model relies on different ML regressors (e.g., random forts, gradient booting) as an alternative to fitted empirical equations used in the ASTM method. Due to the potential of ML-regressors to reflect complex and interdependent relations, the estimation of jet fuel properties can be germane to the capacities of ML models, especially since the By using the same set of input/output variables as the ASTM method to train and test the ML models, the adaption approach focuses on the capacity of the ML models to represent a numerically more efficient and generalized solution for the underlying estimation problem. Such a solution would be less susceptible, e.g., to variance not represented by the training data set. Similar to the evaluation approach, the sensitivity of the resulting models, as well as the correlation to the fuel’s composition, are evaluated.

Finally, the evaluation of the bulk property-based ML-model simplification is based on the formulation of potential ML models that use simple-to-measure parameters, which would, from an application point of view, reduce operational overhead. The approach focuses on models that use only the bulk properties (e.g., density, viscosity, Refractive index, Electrical Conductivity, etc.) and their capacity to model the required quantity, viz. NHC. Within this scope, different sets of ML methods (e.g., random forts, gradient booting) and input parameters are studied. Similar to the other approaches, the sensitivity of the resulting models and the correlation to the fuel’s composition are evaluated.

In conclusion, the comparison of error baselines and sensitivity analysis of the three approaches is intended to bolster the interpretability of ML models and help mitigate the “black box” effect of models. Depending on the established sensitivity of these models, different sets of modeling approaches and input parameters can be considered to improve the estimation accuracy. Initial results are presented in the attached figure comparing ASTM D3338 and the two ML-Modeling approaches for conventional fuels. The results show that the ML-based adaptation model can match the results of ASTM D3338 when tested on the same data set. Furthermore, the simplified ML model based solely on bulk properties (viz. density at 15°C, reflective index, and viscosity at -20°C) has been found to give promising results (root mean square error = 0.002 [MJ/kg] and max relative error < 0.5%). The results suggest that bulk properties and ML-based models can be used to resolve complex quantities such as NHC for well-defined aviation fuels.

Alqaheem and Riazi (2017): Flash Points of Hydrocarbons and Petroleum Products: Prediction and Evaluation of Methods [https://doi.org/10.1021/acs.energyfuels.6b02669]

Edwards (2023): Jet Fuel Properties, United States Air Force Research Laboratory [AFRL-RQ-WP-TR-2020-0017]

Hall Et al. (2021): Predictive Capability Assessment of Probabilistic Machine Learning Models for Density Prediction of Conventional and Synthetic Jet Fuels [https://doi.org/10.1021/acs.energyfuels.0c03779]

Hall Et al. (2022): Probabilistic Mean Quantitative Structure–Property Relationship Modeling of Jet Fuel Properties [https://doi.org/10.1021/acs.energyfuels.1c03334]

Heyne Et al. (2021): Sustainable aviation fuel prescreening tools and procedures [https://doi.org/10.1016/j.fuel.2020.120004]

Vozka, Et al. (2018): Impact of HEFA Feedstocks on Fuel Composition and Properties in Blends with Jet A [https://doi.org/10.1021/acs.energyfuels.8b02787]

ASTM D976: Standard Test Method for Calculated Cetane Index of Distillate Fuels

ASTM D1319: Standard Test Method for Hydrocarbon Types in Liquid Petroleum Products by Fluorescent Indicator AdsorptionASTM D1655: Standard Specification for Aviation Turbine Fuels

ASTM D2386: Standard Test Method for Freezing Point of Aviation Fuels

ASTM D3338: Standard Test Method for Estimation of Net Heat of Combustion of Aviation Fuels

ASTM D3343: Standard Test Method for Estimation of Hydrogen Content of Aviation Fuels Hydrogen content

ASTM D4737: Standard Test Method for Calculated Cetane Index by Four Variable Equation

ASTM D4809: Standard Test Method for Heat of Combustion of Liquid Hydrocarbon Fuels by Bomb Calorimeter (Precision Method)

ASTM D7566: Standard Specification for Aviation Turbine Fuel Containing Synthesized Hydrocarbons