O+P Fluidtechnik 11-12/2024

DICHTUNGEN FORSCHUNG UND

DICHTUNGEN FORSCHUNG UND ENTWICKLUNG PEER REVIEWED PRESENTATION HELD AT ISC 2024 PREDICTING COMPATIBILITY OF SEALING MATERIAL WITH BIO-HYBRID FUELS: DEVELOPMENT AND COMPARISON OF MACHINE LEARNING METHODS Bio-hybrid fuels, derived from sustainable raw materials and green energies, offer a promising alternative to conventional fuels made of mineral oil. Within the cluster of excellence “The Fuel Science Center (FSC)” at RWTH Aachen, bio-hybrid fuels are investigated on a holistic level, including an examination of their compatibility with sealing materials. Previous time-consuming experiments revealed that many biohybrid fuels show poor material compatibility with elastomer sealing materials (e.g. NBR & FKM) leading to issues such as volume expansion, hardness alteration, or chemical reactions upon immersion. These incompatibilities could result in catastrophic failures during practical applications. Due to the high number of possible fuels, fuel-mixtures and therefore fuel/seal combinations a solely empirical approach is impractical. Consequently, this work presents a hands-on framework of machine learning (ML) regression models that predicts material compatibility based on experimental results and selected fluid properties, thereby enabling the preselection of possible fuel/seal combinations for subsequent immersion tests. 1 MOTIVATION AND INTRODUCTION In terms of global climate change, CO 2 -neutral forms of energy from sources other than fossil feedstock are of immense importance to all of society. One possibility for providing climate-neutral energy for automotive applications is the use of bio-hybrid fuels, which are produced either based on biological resources or with the use of other carbon sources and renewable energy. In order to select fuels from the almost infinite number of possible molecules that meet environmental, economical, and technical requirements, the Cluster of Excellence „The Fuel Science Center (FSC)“ is developing methods to identify optimal fuel candidates. One advantage of bio-hybrid fuels over other alternative energy sources, such as fuel cells or electrical energy, is the use of existing infrastructure. If bio-hybrid fuels are used as so-called „drop- T1 Descriptor: Density Molar mass Molar volume Boiling point Enthalpy of vaporization Vapor pressure Oxygen content Ring structure logP ow Overview of selected describing parameters Source: Measured and from literature From literature Engineered from density and molar mass From literature From literature (at standard conditions) From literature (at 20°C) Calculated from molar structure From molecular structure From literature Reference: [23, 24] [23, 24] - [23, 24] [23, 24] [23, 24] [23, 24] [23, 24] [25] Data type and unit: Numerical value kg/m³ Numerical value g/mol Numerical value m 3 / mol Numerical value °C Numerical value kJ/mol Numerical value kPa Numerical value % Either 1 or 0- Numerical value - in fuels“ in existing combustion engines, a worldwide network of filling stations and production facilities in the fuel industry can be utilized. A mandatory prerequisite for the unrestricted use of “drop-in fuels” developed in the FSC is material compatibility with all system components. Static and dynamic seals are particularly at risk. Failure of seals can have far-reaching consequences as has been shown in the past with test bench failures caused by damaged seals. Previous studies and immersion tests have shown that many bio-hybrid fuels have poor material compatibility with conventional elastomer sealing materials, leading to an increase in elastomer volume of more than 200 % [1, 2]. Such large amounts of swelling lead to immediate failure in real technical applications. In addition to swelling, other wear mechanisms such as changes in hardness or chemical reactions were observed [1, 2]. Based on these findings, material compatibility has become a important design parameter for bio-hybrid fuels within the FSC and therefore is addressed in this study. Given the vast array of fuel candidates and blends, along with corresponding fuel/seal combinations, conducting manual immersion tests becomes impractical. In addressing this challenge, an automated test rig was developed, facilitating the simultaneous execution of 90 immersion tests. However, despite this advancement, achieving a comprehensive examination of all potential fuel/seal combinations remains unfeasible. Thus, there arises the necessity for a strategic selection of combinations to investigate further. Machine learning (ML) has proven to be successful in the area of fluid power applications, like fault detection [3] and databased condition monitoring [4]. Furthermore, several studies have been conducted on the discovery of new materials and prediction of fuel properties with ML algorithms [5, 6]. The ability of ML algorithms to discover patterns and trends in a vast amount of 28 O+P Fluidtechnik 2024/11-12 www.oup-fluidtechnik.de

DICHTUNGEN data represents a powerful tool for the identification of material compatibility F1 and the prediction of compatibility of untested combinations, resulting in a narrowed-down fuel/seal configuration. (a) Dataset Hereby, a pairing is valid if the hardness change and sealing material volume after the immersion test are close to the values (b) Training set of immersion tests with conventional fuels, such as diesel or gasoline. The configurations Fit Model are not further considered if the values deviate too much. In the following study, the process of conducting immersion tests, selecting suitable parameters, and preprocessing the resulting dataset is presented. Furthermore, the setup of the ML framework, model training, and evaluation and subsequently the comparison of different regression models are shown. 2 METHODOLOGY In the following section the investigated bio-hybrid fuels and regression methods are introduced. 2.1 INVESTIGATED BIO HYBRID FUELS Different promising drop-in fuel candidates have emerged out of the FSC. Six of them are included in this work. Additionally, diesel in accordance with DIN EN 590 and gasoline according to DIN EN 228 are used as reference fuels [7, 8]. In the following, the composition of the bio-hybrid fuels is shortly introduced. The first representative for bio-hybrid diesel fuels is OME 3-5. OME 3-5 is a mixture of polyoxymethylene dimethyl ethers with chain lengths between 3 and 5 carbon atoms. It is intended to be produced based on methanol and formaldehyde by using low General process of regression tasks T2 Trained Model Comparison of the investigated regression models [14] Regressionmodel : Linear Model Nonlinear Models Linear Regression (polynomial Regression) Lasso Regression Stochastic gradient decent Multi-layer Perceptron Training set (c) Testing set Testing set Trained Model Prediction Actual values Score Workingprinciple: Fits a linear (polynomial) function to the data by minimizing the difference between the true and predicted values, adjusting coefficients to find the best - fitting line (or curve) Additional to linear regression, the absolute size of coefficients is penalized using L1 regularization Minimizes the loss function iteratively by randomly selecting mini batches of data and updating model parameters in the direction opposite to the computed gradient Data is propagated through (multiple) layers of neurons, with each layer performing transformations on the input, ultimately predicting the output by minimizing the loss function PREVIOUS STUDYS SHOWED POOR MATERIAL COMPATIBILITY carbon electricity [9]. The two HyFit fuels can be produced using Fischer-Tropsch process and consist of different shares of n-alcohols and n-alkanes. HyFit1 and HyFit2 differ primarily in their accumulated alcohol content of 15 and 40 % respectively [10]. The third bio-hybrid diesel fuel is a mixture of different methyl ketones which is derived from a microbiological process [2]. Representatives of bio-hybrid Otto fuels are KEAA and EBCC blends which are mixtures of different ketones, esters, alcohols, and alkanes [11]. Most of the fuel candidates mentioned above are mixtures of pure liquids. Some pure liquids have already been tested by the FSC for their material compatibility and are therefore included in this study [2, 12]. This has two advantages. Firstly, the experimental database is expanded, making predictions by ML regression models more significant. Secondly, the foundation is provided to further investigate the material compatibility of mixtures or fuel blends and to understand the underlying patterns. 2.2 REGRESSION MODELS The field of machine learning focuses on finding patterns in data using computer algorithms. Its main purpose is to utilize these patterns to perform various tasks, such as data classification or estimation of continuous values based on the given data [13]. Since the aim of this study is to predict numerical values based on a dataset that consists of both input and their corresponding SVR Decision Tree Regressor Aims to find the hyperplane that best fits the data by maximizing the margin between the data points and the hyperplane and minimizing the loss function Constructs a tree-like structure where data is split into feature subsets, optimizing splits to minimize the variance of predicted values within each subset, resulting in a piecewise constant approximation of the output variable target (output) variables, a supervised learning (SL) approach is selected. In the case of SL, the output is the true value, called a label, and the model should be tuned in such a way that it predicts the particular label for the provided input set. The inputs of the ML models are the fuel properties, which are listed in Table 1. Their selection process is elaborated further in section 3.2. The provided labels are the volume amount and hardness change of the elastomer specimen. This setup represents a multi-output regression problem, which can be solved by a variety of SL algorithms, listed in Table 2. This work aims to implement a framework, which integrates these algorithms and could be extended and validated with currently available and further collected data. The general process for solving regression tasks consists of three steps that are denoted from a) to c) in Figure 1. The first step is splitting up the dataset (a). In a second step, the model is trained (b) and lastly, the performance of the model is evaluated (c). www.oup-fluidtechnik.de O+P Fluidtechnik 2024/11-12 29