Citation: Ricci, F.; Petrucci, L.; Mariani, F. Using a Machine Learning Approach to Evaluate the NOx Emissions in a Spark-Ignition Optical Engine. Information 2023, 14, 224. https://doi.org/10.3390/info14040224 Academic Editor: Libing Wu Received: 20 February 2023 Revised: 21 March 2023 Accepted: 4 April 2023 Published: 6 April 2023 Copyright: © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). information Article Using a Machine Learning Approach to Evaluate the NOx Emissions in a Spark-Ignition Optical Engine Federico Ricci, Luca Petrucci * and Francesco Mariani Engineering Department, University of Perugia, Via Goffredo Duranti, 93, 06125 Perugia, Italy * Correspondence: luca.petrucci89@gmail.com Abstract: Currently, machine learning (ML) technologies are widely employed in the automotive field for determining physical quantities thanks to their ability to ensure lower computational costs and faster operations than traditional methods. Within this context, the present work shows the outcomes of forecasting activities on the prediction of pollutant emissions from engines using an artificial neural network technique. Tests on an optical access engine were conducted under lean mixture conditions, which is the direction in which automotive research is developing to meet the ever-stricter regulations on pollutant emissions. A NARX architecture was utilized to estimate the engine’s nitrogen oxide emissions starting from in-cylinder pressure data and images of the flame front evolution recorded by a high-speed camera and elaborated through a Mask R-CNN technique. Based on the obtained results, the methodology’s applicability to real situations, such as metal engines, was assessed using a sensitivity analysis presented in the second part of the work, which helped identify and quantify the most important input parameters for the nitrogen oxide forecast. Keywords: machine learning; NARX; Mask R-CNN; emissions; engine 1. Introduction Increasingly stringent pollutant emission standards and the fuel economy require- ments put high demands on research into the efficiency of internal combustion engines (ICEs) [1,2]. OEMs are currently developing innovative strategies for future high-efficiency engines able to address this challenge, such as engine boosting and downsizing [3], low- temperature combustions (LTCs) [4], water injection [5,6] and lean [7,8] and/or EGR- diluted mixtures [9]. However, during the engine calibration process, the optimization of the efficiency and emissions requires engine parameters to be adjusted through extensive activities [10]. Moreover, the fine control of important operation variables can sometimes be hard to reach due to the inherent limitations of the measuring instruments [11,12]. Currently, machine learning (ML) approaches are widely used to solve problems in the automotive field, thanks to their ability to identify the intrinsic relationship between the input parameters and the engine response [13,14], with lower computational costs and faster operations than traditional methods [15,16]. ML algorithms demonstrated excellent results in predicting engine parameters such as pressure [17], fuel consumption [18], exhaust gas temperature [19], power [20] and emissions [21]. Considering the latter, Yaopeng Li et al. [21] employed an artificial neural network (ANN) with a genetic algorithm (GA) to optimize a direct dual fuel stratification (DDFS) strategy, starting from a numerical model of a light-duty diesel engine based on the General Motors 1.9 L platform. The optimized parameters (i.e., in-cylinder pressure and tempera- ture, EGR rate, injection timing of fuels) were validated across a wide operating range. The performance was compared to that of a GA-CFD (computational fluid dynamics) approach. The ANN–GA method allowed improved fuel efficiency and lower nitrogen oxide (NOx) emissions to be obtained with lower computational time (over 75% of computational time saving). Samrendra K. Singh et al. [22] combined a genetic algorithm with a machine Information 2023, 14, 224. https://doi.org/10.3390/info14040224 https://www.mdpi.com/journal/information