Browse Topic: Artificial intelligence (AI)

Items (2,399)

Machine Learning-Based Gear Defect Detection in Electric Vehicle Drive Units Using End-of-Line Vibration Analysis

2026-01-0637To be published on 04/07/2026

The final assembly of electric vehicle (EV) drive units includes an essential End-of-Line (EOL) test to ensure both component integrity and Noise, Vibration, and Harshness (NVH) quality. This screening process, which uses dynamometers to measure vibration signals, is critical for identifying defects before a drive unit is installed in a vehicle. A significant source of failure during this test is gear defects, which can arise from manufacturing or handling issues. Traditional EOL testing methods rely on time-domain analysis and the impulsiveness of vibration signatures to detect these defects, a technique with inherent limitations in accuracy. This paper introduces and evaluates a novel approach using Machine Learning (ML) to analyze vibration signals for improved gear defect detection. We discuss the methodologies of both the traditional time-domain and the proposed ML-based techniques. Finally, we provide a comprehensive comparison of their respective efficiency and accuracy

Arvanitis, Anastasios, Michaloliakos, Anargyros

Modular Smart Corner Systems for Next-Generation Electric Vehicle Architecture

2026-01-0639To be published on 04/07/2026

The transition to software-defined vehicles (SDVs) necessitates a paradigm shift in both control strategies and vehicle architecture. The EU-funded SmartCorners project addresses this challenge by developing integrated smart corner systems (SCS) that combine in-wheel motor propulsion, brake-by-wire, active suspension, and steer-by-wire functionality in a single module. These SCS are embedded within a highly adaptable skateboard chassis architecture, enabling scalable deployment across diverse vehicle platforms. The central approach of this paper is the utilization of artificial intelligence and machine learning to implement multi-layer, data-driven control strategies, facilitating real-time actuation, fault mitigation, and a user-centric vehicle architecture. The SmartCorners project targets to demonstrate significant advancements, including improvement in real-world driving range, reduction of component and system costs, and development time drop of the EV systems enabled by digital

Ratz, Florian, Armengaud, Eric, Formento, Cecilia, Moscone, Giulia, Sorrentino, Gennaro, Bisciaio, Giorgio, Sorniotti, Aldo, Amati, Nicola, Braun, Daniel, Deibler, Bernd, Boxberger, Valerius, Sottile, Salvatore, Ivanov, Valentin, Fuse, Hiroyuki, Kompara, Tomaž

Curvature-Based End-to-End Autonomous Vehicle Path Planning at Intersections

2026-01-0045To be published on 04/07/2026

Autonomous vehicle navigation requires accurate prediction of driving path curvature to ensure smooth and safe trajectory planning. This paper presents an end-to-end approach to curvature prediction using deep neural networks trained on real-world driving data. We develop a comprehensive neural network architecture that directly maps high-dimensional sensor inputs to curvature predictions, eliminating the need for intermediate feature engineering steps. The model processes multi-modal inputs including vision features, vehicle state parameters, and environmental context through a deep learning pipeline. Our end-to-end approach demonstrates the feasibility of learning complex driving behaviors directly from raw sensor data, providing a more robust and generalization solution compared to traditional rule-based methods. The neural network architecture incorporates dropout regularization and adaptive learning rate scheduling to ensure stable training and optimal performance. Results show

Hajnorouzali, Yasaman, Wang, Hanchen, Li, Taozhe, Burch, Collin, Lee, Victoria, Tan, Lin, Arjmandzadeh, Ziba, Xu, Bin

Is Implicit Knowledge Enough for LLMs? A RAG Approach for Tree-based Structures

2026-01-0111To be published on 04/07/2026

Large Language Models (LLMs) are effective at generating responses using contextual information, which is especially useful when working with structured data like code. Retrieval-Augmented Generation (RAG) enhances this capability by retrieving relevant documents to supplement the model’s context. However, the best way to represent retrieved knowledge—particularly for hierarchical structures such as trees—remains underexplored. In this work, we introduce a novel method to linearize tree-structured knowledge, making it suitable for storage in a knowledge base and direct use with RAG. We compare this approach to applying RAG on raw, unstructured code, evaluating both the accuracy and quality of the generated responses. Our results show that while response quality is comparable between the two methods, our linearized approach reduces the number of vector documents by almost 4x. Without any loss of generality, we applied and benchmarked our method to a specific automotive use case, that is

Gupte, Mihir, Giusto, Paolo, S, Ramesh

Artificial Intelligence in Automotive Stamping –the latest industrial applications and future directions

2026-01-0179To be published on 04/07/2026

A review of the applications of Artificial Intelligence (AI) in automotive stamping is presented. The focus is on recent AI implementations within the automotive industry. Through this review, the authors aim to capture the current momentum of AI in automotive stamping. The article begins with an overview of the importance and challenges of stamping in the automotive sector, followed by a discussion of key AI technologies applied in this domain. Several industrial applications are then introduced, categorized by their specific use cases. Finally, strategic challenges and future directions are discussed.

Sheng, ZiQiang, Huang, Lu, Asimba, Brian, McCarty, Eric, whaley, Jason, cabral, Kleber, Osegueda, Mario, Huang, Xiaosong, Erol, Baris

Energy-Aware Traffic Signal Optimization for Sustainable Urban Networks Using Valid Action Policies

2026-01-0462To be published on 04/07/2026

Optimizing signal control across urban networks is crucial for reducing energy consumption and improving traffic efficiency. Advances in Vehicle-to-Everything (V2X) technologies have enabled deep reinforcement learning (DRL) to demonstrate significant promise in traffic signal control. However, many existing studies focus primarily on agent performance while neglecting practical operational constraints, which can lead to fairness issues in traffic control. Moreover, prior work often treats energy optimization as a by-product of efficiency. To address these gaps, this study proposes a multi-agent DRL framework to mitigate peak-hour congestion and reduce excessive emissions in urban networks. Building on this framework, a variant of Independent Proximal Policy Optimization (IPPO), termed Global State Independent PPO (GS-IPPO), is developed to enhance IPPO's coordination capability. To further improve practical applicability and fairness, this study incorporates constraints imposed by the

Qian, Wangrui, Jiang, Shuyan, Qi, Hao, Ou, Shiqi(Shawn)

Peak SCR Efficiency for Ultra-Low NOx Using Electrically Heated Mixer for EU-VII and China-VII

2026-01-0363To be published on 04/07/2026

Achieving ultra-low NOx emissions remains a major challenge in diesel emission control worldwide, especially as increasingly stringent regulations are introduced globally. Selective Catalytic Reduction (SCR), the leading NOx reduction technology in diesel systems, performs best when both heat and ammonia are available. At the same time, any proposed solution must also be cost-effective and economically viable. In this work, we present a low-cost Electrically Heated Mixer (EHM). EHM not only significantly reduces urea deposit formation, thereby lowering warranty costs and mitigating failure modes, but also enhances SCR efficiency, especially in challenging low-load conditions having low exhaust temperature. EHM is also ideal for rapid heat-up and urea injection during engine cold-start conditions ( <100 °C) enabling swift NH₃ formation and storage in the SCR catalyst. Additionally, it enables early urea injection (~ 130 - 150 °C), improving SCR efficiency in low temperatures below 200

Masoudi, Mansour, Poliakov, Nick

A Frequency-Aware Reinforcement Learning Framework for Suspension Control with Integrated Temporal Features and Expert Guidance

2026-01-0209To be published on 04/07/2026

Active suspension systems play a crucial role in improving vehicle ride comfort and handling stability. However, most existing studies focus on the low-frequency range below 20 Hz, leaving the suppression of high-frequency vibrations within 50–500 Hz largely unexplored, even though these vibrations strongly affect in-cabin noise and ride quality. To address this gap, this study introduces a quarter-car suspension model incorporating both bushing dynamics and a rigid-ring tire within a reinforcement learning (RL) framework. A major challenge for RL-based suspension control is its degradation in high-frequency performance. To overcome this issue, we design an innovative training framework that integrates multiple synergistic strategies. First, frequency-domain rewards are incorporated as auxiliary signals to explicitly guide policy optimization in the high-frequency band. Second, long short-term memory (LSTM) networks are embedded in both the Actor and Critic to capture the sequential

zhu, Zhehui, Zhang, Lijun, Meng, Dejian, Hu, Xingyu

In-situ Decoupling and Stiffness Optimization of Resiliently Coupled Assemblies Using FRF-Based Substructuring

2026-01-0227To be published on 04/07/2026

Road shake remains one of the most critical NVH challenges in light- and heavy-duty body-on-frame trucks, directly shaping ride comfort and perceived vehicle quality. At the heart of this issue are the body mounts, which govern the transfer of vibration between frame and body. Optimizing these mounts for improved road shake performance is therefore a key objective in NVH engineering. Traditionally, such optimization has relied on finite element analysis (FEA). While highly accurate, FEA—especially when performed with commercial solvers such as NASTRAN—demands enormous computational resources, with complete simulations often requiring hundreds of hours. Each design iteration multiplies this cost, restricting the pace and efficiency of NVH development. Recent methods such as response surface modeling (RSM) and machine learning (ML) have aimed to reduce the burden, yet they remain dependent on large libraries of pre-computed FEA data, inheriting much of the same expense and inflexibility

Haider, Syed, Abbas, Ahmad, Jahangir, Yawar, Maddali, Ramakanth

From LLM to Deep Learning: Efficient Simulation of Last-Mile Energy Behavior in Campus Communities

2026-01-0461To be published on 04/07/2026

Communities are critical nodes in the urban energy network, integrating energy, transportation, and building systems to enhance overall energy efficiency. Accurate management of these systems depends on characterizing individual energy-demand behaviors. However, traditional last-mile behavioral models often fail to capture the complex and non-linear nature of human decision-making, creating a need for advanced simulation techniques. Although agent-based simulations powered by Large Language Models (LLMs) have demonstrated considerable efficacy across diverse domains, their substantial computational demands—specifically in terms of token consumption—pose pronounced constraints in large-scale simulations involving tens of thousands of agents, considerably limiting their practical applicability. To address this challenge, a deep learning-based approach is proposed for single-step prediction of last-mile behavior. A hybrid architecture, consisting of an MLP encoder and a Transformer

Yang, Zhifeng, Chen, Yongjian, Ou, Shiqi(Shawn)

Study on the Regulation of the Three-Phase Microenvironment in Low-Platinum MEA

2026-01-0443To be published on 04/07/2026

With the growth of energy demand, fuel cells as efficient and clean energy devices, have attracted increasing attention. However, the high cost of membrane electrode assembly (MEA) restricts their large-scale application. Therefore, reducing the platinum usage and improving performance have become key research point. In this work, MEA was prepared and excellent performance of 1.52 W·cm-2 was achieved at a low platinum loading. The influence of different ionomer/carbon (I/C) ratio on the performance of fuel cells was systematically investigated. It was found that the performance of the MEA was the highest when the I/C ratio is 0.6. Quantifying hydrophilic and hydrophobic characteristics of catalyst layers with varying ionomer contents revealed that the proton conduction efficiency is optimal when the I/C ratio is 0.6. This balance established efficient proton conduction pathways, from the results of proton conduction impedance testing. SEM analysis demonstrated that pore structure

Li, Xin, Cai, Xin, Lin, Rui

Adaptive Economic Model Predictive Control for Engine Emissions Management with Compensation for Performance Drift

2026-01-0287To be published on 04/07/2026

This paper addresses the changes in engine emissions due to in-use component changes through the synergistic application of predictive control, machine learning, and onboard adaptation. In particular, we consider an adaptive economic Model Predictive Control (eMPC) strategy to mitigate the effects of performance drift on Nitrogen Oxides (NOx) and Soot emissions from compression ignition (diesel) engines. A performance drift block, which applies a multiplier and offset to nominal emissions, is integrated with a high-fidelity Neural Network (NN) plant model to simulate these characteristic changes. To counteract variability, two online adaptation methods are integrated within the eMPC framework: One is based on Recursive Least Squares (RLS) and another on a continuously updated online NN. The proposed control architecture is validated through simulations over standard transient cycles. Results demonstrate that while the rate-based eMPC possesses inherent robustness to performance drift

ZHANG, JIADI, Li, Xiao, Kolmanovsky, Ilya, Tsutsumi, Munecika, Nakada, Hayato

CAE Acceleration With Machine Learning For Results Prediction

2026-01-0488To be published on 04/07/2026

Industries are following a tedious product development cycle for developing their product. In product development major steps includes design ideas, Drawings, CAD, CAE, Testing and design improvement cycle. This is a monotonous process and takes time which impacts on its time to deliver product and cost on development. Now a days industries are fast growing and targeting to reduce development cycle time and cost. AIML is impacting almost all areas in the industry and significantly reducing efforts time and cost. To make use of AIML in CAE, Altair Physics AI is an effective tool. To ensure the design of product traditional way is to develop a CAD of the product, develop, perform CAE and analyze performance. If we consider CAE procedure it is time consuming process which includes FEA model build, applying boundary conditions, running simulation and analyzing results which could take minutes to hours. By using ML with Physics AI we can make predictions on new design of the product in

Dangare, Anand Manohar, Kulkarni, Mandar

AI LabelMate: A Context-Aware Annotation Agent for Reducing Semantic Fragmentation

2026-01-0261To be published on 04/07/2026

Robust perception systems for autonomous vehicles depend critically on high-quality labeled data, especially for off-road and unstructured environments. However, perception model performance is often degraded by 'data chaos' from limitations in automated segmentation. Foundation models like SAM2, while powerful, usually generate masks based on low-level visual cues like color and texture gradients. In complex off-road scenes, this leads to semantic fragmentation. A single object, like a moss-covered log, can be split into not only dozens of segments for its bark and moss, but also hundreds of smaller, meaningless patches based on minor color variations. This paper introduces a context-aware annotation agent to resolve this issue. Our workflow integrates a vision-language model (Florence-2) for scene understanding with a segmentation model (SAM2) for mask generation. Instead of segmenting indiscriminately, our agent leverages Florence-2 to comprehend the image holistically, localizing

Patil, Ashish, Mikulski, Dariusz, Mwakalonge, Judith, Jia, Yunyi

HaloBus: Edge Computing‑Enabled Real‑Time Boarding and Exit Detection for Enhanced Transportation Safety Using Lightweight AI

2026-01-0555To be published on 04/07/2026

This paper proposes HaloBus, an innovative, edge‑computing solution designed to mitigate this risk by detecting student boarding and exiting in real time using lightweight AI based methods. A persistent challenge in elementary school transportation is the issue of missing students after they exit their buses, which disproportionately impacts low-income households. Current safety systems are forced upon the individual households with their own methods. Common methods include applications on a personal device or a small tracker. However, not everyone can afford these options, and every parent deserves to feel like their child is safe. That is why HaloBus was invented. The system employs YOLOv5u—an Ultralytics‑enhanced, anchor‑free, split‑head architecture that offers a superior accuracy–speed tradeoff. By providing real-time, on-device alerts, HaloBus enables immediate intervention to prevent a student from being left behind, thereby shifting the focus from reactive post-incident

Getz, Grayson, Zadeh, Mehrdad, TAN, Teik-Khoon

Computational Modeling and Optimization of Shape Memory Polymer-Based Energy Absorbers

2026-01-0573To be published on 04/07/2026

Shape memory polymers (SMPs) provide tunable thermomechanical properties and enable the design of recoverable crash structures for automotive applications. This paper introduces a computational framework for the design and optimization of SMP-based crash absorbers with periodic auxetic microstructures. A finite element (FE) model is developed and validated against experimental data reported in the literature. The model is used to simulate crushing and recovery behavior under different temperature conditions. A parametric study is performed by varying key microstructural features, including wall thickness, cell size, and cell shape. Structural performance is evaluated in terms of specific energy absorption (SEA), peak force, and recoverability under axial crush loading. To efficiently explore the high-dimensional design space, surrogate models based on machine learning are constructed, and multi-objective optimization is carried out to identify Pareto-optimal designs that balance

Zhu, Yingbo, Zhu, Feng, Deb, Anindya

Extreme Scale Automotive Aerodynamic Simulations

2026-01-0597To be published on 04/07/2026

Achieving an optimal balance between simulation accuracy and computational efficiency remains a central challenge in automotive aerodynamics. While the adoption of AI and machine learning (ML) methods in vehicle development is expected to grow significantly, the demand for highly scalable, computationally efficient, and accurate computational fluid dynamics (CFD) methods persists. The emergence of GPU technology presents new opportunities to deliver cost-effective, high-fidelity, scale-resolving simulations to industrial users. A comprehensive evaluation of Simcenter STAR-CCM+’s parallel scalability and accuracy across extensive CPU and GPU resources was executed on the Frontier supercomputer at Oak Ridge National Laboratory. Steady-state and transient aerodynamic scalability simulations were executed using the DrivAer notchback vehicle configuration. Simulation accuracy was evaluated through transient simulations employing the SST-DDES turbulence model on an initial mesh of 200

Larsson, Torbjörn, Grover, Ronald O., Landi, Simone, Altmann, Peter, McManus, Liam, Dowding, Steven

Development of Windshield Wetting Level Assessment Using Computer Vision

2026-01-0524To be published on 04/07/2026

This paper introduces a novel methodology for evaluating windshield wetting (rain) levels by leveraging sensor fusion approach combined with advanced computer vision applications. The proposed method aims to enhance the traditional wetting detection systems by providing an alternative or granular classification of various wetting conditions. We will discuss the challenges encountered during the development and validation phases, including variability in rain intensity, droplet dynamics, and environmental factors. Furthermore, the evaluation process of critical parameters influencing wetting detection accuracy will be presented, detailing the metrics used for performance assessment. Initial results from experimental evaluations demonstrate the efficacy of the new method in diverse wetting scenarios, and an in-depth analysis of challenging edge cases, such as mist, glare, and partial wetting, will be provided. This research contributes to the driver experience.

Schroeter, Robert A., Ghannam, Mahmoud, Shaik, Faizan

Virtual Sensor Development for Real-Time Estimation of Transient NOx Emissions for Internal Combustion Engine

2026-01-0370To be published on 04/07/2026

Accurately measuring NOx emissions under transient engine conditions is becoming increasingly important with upcoming Euro 7 and EPA 2027 regulations. Traditional physical sensors often struggle with cost and response time, especially with aging of sensors in dynamic operation. This paper introduces a machine-learning–based virtual NOx sensor that can provide real-time emission estimates while reducing reliance on hardware sensors. The approach uses multiple machine-learning methods (Random Forest, Bootstrap Aggregating, Adaptive Boosting, Gradient Boosting, Extreme Gradient Boosting) and selected best one to establish correlations between engine operating parameters, measured steady-state data, and transient duty cycle NOx emissions. Validation across different duty cycles has shown strong alignment with physical sensor readings, with R² values above 0.9995 for known data sets and above 0.9534 for unseen conditions. The model needs to be trained with larger training samples to further

Kumar, Chandan, Dahodwala, Mufaddel, Thawrani, Kiran

Development of a System for Regression Model Enhancement Utilizing Shape Generation AI

2026-01-0499To be published on 04/07/2026

This study proposes a method to enhance regression models by shape generation AI. The approach focuses on automatically identifying regions within the design space where the model’s prediction accuracy is low. Once these regions are identified, new and diverse sample shapes are automatically generated by the shape generation AI and incorporated into the training dataset. The regression model is then retrained to improve its performance. By iteratively repeating this cycle of exploration, shape (FE mesh) generation, and model updating, the model’s reliability and accuracy across the entire design space are progressively enhanced. This method addresses data sparsity issues common in complex design tasks and enables better generalization to underrepresented regions. The effectiveness of the proposed system was demonstrated through a case study involving hood outer panels in automotive design. The results showed that adding AI-generated shapes improved prediction accuracy, particularly in

Taniguchi, Mashio

Neural Surrogates and Reinforcement Learning for Heavy-Duty Diesel Torque Control: A Data-Driven Alternative to PID

2026-01-0199To be published on 04/07/2026

Accurately reproducing the torque trace of regulatory drive cycles is central to heavy-duty diesel emissions certification and repeatable development testing. Conventional controllers—PI/PID with gain scheduling and feedforward mapping—can meet requirements but demand extensive hand-tuning and struggle with strong nonlinearities and time-varying delay. Building on prior work that quantified the limits and benefits of gain-scheduled and feedforward strategies under the FTP transient cycle, this work investigates a fully data-driven control framework. First, a recurrent neural network (LSTM) torque surrogate is trained on dynamometer data to capture engine dynamics from readily available signals (e.g., engine speed, commanded pedal/torque, recent torque history) and actuator constraints. Second, this learned “world model” is used to train a reinforcement learning (RL) policy that outputs pedal commands to minimize torque-tracking error while regularizing control effort and respecting

Cook, James, Puzinauskas, Paulius, Bittle, Joshua, Hall, Spencer

Evaluating the Impact of Drag Coefficient Changes on Driving Energy Using a Probability Distribution Model of Yaw Angle Based on Empirical Data from North American Highways

2026-01-0615To be published on 04/07/2026

This study estimates the impact on driving energy of differences in aerodynamic characteristics for yaw angle from natural wind during North American Highway mode driving. A previous study[1] clarified the potential to estimate the fuel consumption impact of natural wind by integrating the CD yaw characteristics and yaw angle occurrence frequency. The natural wind was measured on a vehicle while driving a representative North American Highway test course[2]. The driving energy is predicted from the obtained yaw frequency and CD yaw sweep data in a wind tunnel. Measurements were conducted every weekday for 8 hours in 2023, covering 70% of the traffic volume. The validity of the measurement period was evaluated by the deviation from the annual average of wind direction and speed. Since yaw frequency varies depending on the road environment, it is necessary to weight the road environment type frequency when calculating the driving energy. The frequency was calculated using machine

Onishi, Yasuyuki, Nucera, Fortunato, Nichols, Larry, Metka, Matt

Enhancing Vehicle DMU Processes Through Artificial Intelligence (AI) (Revolutionizing Packaging Design)

2026-01-0102To be published on 04/07/2026

Abstract- Digital Mock-Up (DMU) is a vital process in the automotive industry, enabling teams to visualize, analyze, and optimize vehicle designs before production. Artificial Intelligence (AI) in the Digital Mock-Up (DMU) processes used in vehicle development. DMU is a virtual representation of a vehicle or component that allows for design, analysis, and simulation without the need for physical prototypes. By introducing artificial intelligence into vehicle DMU, manufacturers can leverage advanced analytics, automation, and predictive modelling to streamline workflows and enhance design accuracy. AI in digital mock-up helps optimizing space, resulting in smaller car design. This explores how AI-driven approaches are transforming DMU processes, improving efficiency, and shaping the future of automotive innovation. Key Benefits of AI in Vehicle DMU: - Enhanced Design Analysis: AI-enhanced design analysis enables teams to rapidly assess complex automotive models, detecting issues early

Nyati, Mohit Ashok, pathak cEng, Amit

Transformer-Based Deep Learning Framework for Efficient Prediction of Automotive Wind Noise

2026-01-0222To be published on 04/07/2026

Aerodynamic wind noise is a critical challenge in modern automotive development, particularly with the rise of vehicle electrification and intelligent mobility, where cabin acoustic comfort is a key quality metric. While reliable, traditional methods like wind tunnel experiment and CFD simulations are both costly and time-consuming. To address this, we propose a novel transformer-based framework for rapid and accurate wind noise prediction. Several model improvements, including the physical attention, geometry wave number embedding, hybrid FPS-random down sampling method and frequency separation output heads are properly employed to reduce the GPU mermory cost and improve the prediction accuracy. This framework is pre-trained on a large-scale transient flow field dataset of 1,000 diverse vehicles generated using Delayed Detached Eddy Simulation (DES). From a vehicle's point-cloud coordinates, the model directly predicts the surface pressure spectrum on the driver-side window and the

TANG, Weishao, Liu, Mengxin, QIN, Ling, Duan, Menghua, wang, chengjun

Multi-Physics Topology Optimization and Machine Learning Surrogates for Lightweight Cast Automotive Parts

2026-01-0510To be published on 04/07/2026

This paper proposes an integrated optimization framework for lightweight cast vehicle structures by combining Multi-Physics Topology Optimization (MPTO) with machine learning (ML)-driven surrogate modeling. Conventional structural optimization approaches often rely on simplified load cases—such as 3 g inertial loading, frame torsion, or bending—rather than capturing the true complexity of performance-critical conditions like crashworthiness and fatigue durability. As a result, designs optimized for a single criterion frequently degrade in other performance areas, leading to costly redesigns or over-engineered solutions. The proposed methodology introduces MPTO through four structured steps that progressively build design fidelity. Step 1 establishes a baseline primitive topology optimization workflow using a standardized package checklist, ensuring that the design domain, boundary conditions, constraints, and key manufacturing rules are consistently defined as the foundation for

Kim, Hyosig, Senkowski, Andres, Gona, Kiran, Saroha, Lalit, Boraiah, Mahesh

Leveraging Large Language Models for CAD Automation and Multi-Objective Optimization of Automotive Heat Exchangers

2026-01-0506To be published on 04/07/2026

The design of automotive heat exchangers is a critical challenge, requiring the careful balance of thermal performance, aerodynamic efficiency, structural integrity, and manufacturability. Traditional design processes rely heavily on manual CAD modeling and iterative simulations, which are both time-intensive and computationally demanding. Recent advances in artificial intelligence, particularly in Large Language Models (LLMs) and Transformer-based architectures, open new possibilities for automating these tasks. This work introduces a novel framework that integrates LLMs with CAD kernels to generate parametric heat exchanger geometries directly from natural language descriptions. The models are automatically exported into standard formats for computational fluid dynamics simulations and seamlessly incorporated within a multi-objective optimization pipeline. By coupling natural language-driven CAD automation with advanced optimization techniques, the framework enables rapid exploration

Chaudhari, Prathamesh, Tovar, Andres

Leveraging Machine Learning for Occupant Body Size Estimation via B-Pillar Seatbelt Systems

2026-01-0559To be published on 04/07/2026

Occupant body size in vehicles varies significantly, encompassing differences in height, weight, and overall body composition. Adaptive restraint systems, featuring adjustable parameters such as belt load limiters, steering column load limiters and stroke, seat pan stiffness, and airbag pressure, can offer more equitable protection tailored to individual body sizes. In this study, a test rig modeled after the Volvo XC90 was used to collect data from 47 seated participants. Key seatbelt-related parameters, including D-ring angle, belt payout length, lap belt length, and buckle tension, were measured. These measurements were used to train machine learning models to predict occupant characteristics: height, weight, sitting height. The results demonstrate that low-cost sensors embedded in the seatbelt system can provide sufficiently accurate occupant body size estimations to inform adaptive restraint systems. The prediction errors across both training and test datasets were as follows

Wang, Da, Ahmed, Jawwad, Rowe, Mike, Brase, Dan

Machine Learning based Meta-Analysis of IIHS Moderate Overlap Frontal Crash Test for Rear Seat Occupant Safety Performance

2026-01-0579To be published on 04/07/2026

A machine learning (ML)-based meta-analysis was conducted to evaluate rear seat occupant safety performance in the Insurance Institute for Highway Safety (IIHS) Moderate Overlap Frontal (MOF) 2.0 crash test. ML models were trained on historical IIHS crash test data to predict rear passenger injury metrics using vehicle architecture, restraint system characteristics, crash pulse parameters, and vehicle kinematics as input features. The models demonstrated high predictive accuracy and were subsequently used in a Sobol sensitivity analysis to identify critical design parameters influencing injury outcomes. The analysis revealed that rear passenger injury metrics were most sensitive to restraint system parameters. Specifically, crash pulse magnitude was the dominant factor for head injury metrics, pretensioner activation time for neck tension force, and lap belt force for the Neck Injury Criterion (Nij). For chest-related metrics—sternum deflection, dynamic belt position, and maximum belt

Lalwala, Mitesh, Kim, Wonhee, Furton, Lisa, Song, Jay

Culture Clashes: LLM Support in the Engineering of Safety-Critical Systems

2026-01-0110To be published on 04/07/2026

The proven usefulness of large language models (LLMs) as tools for software development and the recent rapid increase in their capabilities have made it possible and attractive to extend their scope of application to almost all tasks in the engineering of complex and even safety-critical systems. Although their application may be economically attractive, LLMs are prone to errors and the generated engineering artifacts may not meet the stringent quality requirements for safety-critical systems. In our paper, we systematically analyze the various potential uses of LLM in the engineering lifecycle of safety-critical systems and identify associated risks and approaches to risk mitigation. We classify these uses and compare them with approaches to mitigating the risks associated with traditional design automation. In addition, we address cultural and psychological aspects of the trust placed in LLMs as engineering tools and reliance on their results. Our analysis shows that, despite their

Thomas, Carsten, Wagner, Michael

Attention and Intrinsic Curiosity-Enhanced Deep Reinforcement Learning for Path Planning in Dynamic Environments

2026-01-0044To be published on 04/07/2026

Efficient and reliable path planning remains a core challenge for autonomous vehicles operating in dynamic and crowded environments. Although Deep Reinforcement Learning (DRL) has shown considerable potential in autonomous decision-making, it still faces challenges such as insufficient feature extraction, sparse rewards, and low obstacle avoidance efficiency in complex scenarios. To address these issues, this paper proposes an end-to-end path planning framework, PPO-ICM-Attn. Built upon the Proximal Policy Optimization (PPO) algorithm, the framework incorporates a dual-channel attention convolutional neural network module (Attention-CNN) to enhance spatial and semantic understanding of dynamic obstacles, and introduces an Intrinsic Curiosity Module (ICM) to promote active exploration in sparse reward settings. Furthermore, a reactive avoidance reward function based on velocity obstacle theory is designed and embedded to achieve real-time proactive collision avoidance in highly dynamic

Shen, Shiquan, Liu, Jiahao, Chen, Zheng, Li, Zongdian, Zhao, Yuting, Wu, Minggong, Zhao, Jie, Qin, Zongquan, Wang, Yanfeng

Advanced Damping Force Modeling Using Machine Learning for Next-Generation Electric Vehicle Suspensions

2026-01-0586To be published on 04/07/2026

The rapid evolution of electric vehicles (EVs) has necessitated innovative approaches to optimize ride comfort, handling, and overall suspension performance. Unlike conventional internal combustion engine vehicles, EVs introduce unique challenges due to their distinct weight distribution, powertrain dynamics, and noise characteristics. This paper presents an advanced damping force modeling methodology leveraging machine learning (ML) techniques to enhance the suspension design process for next-generation EVs. The study employs data-driven ML algorithms, such as Gradient Boosting, Random Forest, and Neural Networks, to model the nonlinear and frequency-dependent behavior of dampers under various operational conditions. A comprehensive dataset, generated through simulation and experimental testing, captures the effects of road profiles, vehicle dynamics, and damping settings. The proposed ML-based approach predicts damping forces with high accuracy, outperforming traditional analytical

Hazra, Sandip, Tangadpalliwar, Sonali, Khan, Arkadip

Automotive Crash Dynamics Modeling Accelerated with Machine Learning

2026-01-0568To be published on 04/07/2026

Crashworthiness assessment is a critical aspect of automotive design, traditionally relying on high-fidelity finite element (FE) simulations that are computationally expensive and time-consuming. This work presents an exploratory comparative study on developing machine learning-based surrogate models for efficient prediction of structural deformation in crash scenarios using the NVIDIA PhysicsNeMo framework. Given the limited prior work applying machine learning to structural crash dynamics, the primary contribution lies in demonstrating the feasibility and engineering utility of the various modeling approaches explored in this work. We investigate two state-of-the-art neural network architectures for modeling crash dynamics: MeshGraphNet, a graph neural network that is widely employed in physics-based simulations, and Transolver, a transformer-based architecture with a physics-aware attention mechanism designed to maintain linear computational complexity with respect to geometric

Nabian, Mohammad Amin, Chavare, Sudeep, Akhare, Deepak, Ranade, Rishikesh, Cherukuri, Ram, Tadepalli, Srinivas

From CFD Simulations to Machine Learning: A Comparative Study of ML Tools for Aerodynamic Drag Prediction

2026-01-0598To be published on 04/07/2026

This study presents the development of machine learning frameworks for predicting aerodynamic drag coefficients in automotive vehicle designs using data derived from computational fluid dynamics (CFD) simulations. The input data includes both structured tabulated design parameters and unstructured geometric representations in the form of point clouds, enabling a multi-modal approach to aerodynamic modeling. A range of machine learning models was explored, from traditional regression techniques to advanced deep learning architectures such as graph neural networks (GNNs), each tailored to handle specific data formats. These frameworks were designed to integrate seamlessly with CFD workflows, supporting efficient training and evaluation using standard regression metrics. The work aims to establish a foundation for data-driven aerodynamic analysis in automotive design, emphasizing the role of machine learning in enhancing simulation-based engineering processes.

Kumar, Gaurav, Khanna, Susheel

Machine Learning-Based Prediction of Wind-Induced Interior Noise in Ground Vehicles

2026-01-0599To be published on 04/07/2026

In response to increasing customer demand for enhanced passenger comfort and perceived vehicle quality, OEMs in automotive and commercial vehicles are placing significant emphasis on reducing the interior cabin noise. At highway speeds, wind noise is a primary contributor to the overall noise within the vehicle cabin. Conventional approaches to predict vehicle wind noise rely on physical testing, which can only be conducted in the later stages of the design process once a physical prototype is available. Increased adoption of established computational fluid dynamics (CFD) methods has enabled earlier assessment. However, such simulations require several hours to complete, posing a challenge in the context of rapid design iteration cycles. With the growing adoption of artificial intelligence in engineering, machine learning (ML) approaches have been proposed to predict a vehicle’s aerodynamics performance. Nevertheless, development of ML techniques in the context of aeroacoustics

Higgins, John, Fougere, Nicolas, Sondak, David, Senthooran, Sivapalan, Moron, Philippe, Jantzen, Andreas, Bi, Jing, Oancea, Victor

AI-Driven Data Consistency and Relationship Inference System for Agile Component Library Management

2026-01-0109To be published on 04/07/2026

Reliable component libraries are the foundation of the engineering process and the starting point for all intelligence within CAD tools. In practice, however, libraries created and maintained by librarians often contain incomplete, inconsistent, or outdated data. This paper introduces the component data consistency and relationship inference AI system, developed within Amoeba software, which addresses these challenges by improving component library quality. The system uses AI to infer component attributes such as component type, gender, color, material, etc. Moreover, it can identify relationships such as the family a connector is associated with based on its attributes and geometry. The system improves data consistency in areas such as resolving mismatched wire size constraints imposed by the connector and cavity components. It also utilizes computer vision to identify common connector footprints, cavity sizes, and 2D symbol geometries. Deployed within Amoeba software, the system has

Phan, Dung, Horvat, Bryan

AI-Enhanced Functional Safety in ADAS Controllers: Predictive Fault Management under ISO 26262

2026-01-0036To be published on 04/07/2026

Ensuring ISO 26262 functional safety in advanced driver assistance systems (ADAS) is increasingly complex as these platforms adopt artificial intelligence for perception and control. Traditional safety mechanisms are deterministic, but AI introduces non-determinism that challenges verification and validation. This paper presents a predictive fault management framework that enhances functional safety in ADAS controllers using AI-driven models. Real-time vehicle and sensor data are processed through machine learning algorithms that forecast hardware and software faults before they escalate into hazardous conditions. These predictive insights are coupled with ISO 26262 safety mechanisms to enable adaptive diagnostics, fault isolation, and recovery strategies. Hardware-in-the-loop (HIL) validation demonstrates up to a 25% improvement in diagnostic coverage compared to conventional monitoring, while maintaining ASIL-D compliance and real-time performance. By showing how AI can complement

Abdul Karim, Abdul Salam

Ensemble of Experts for Generating Digital Parking Maps from Remote Sensing Images

2026-01-0107To be published on 04/07/2026

Being a vital component of transportation, parking and parking-related research are gaining increasing attention from the intelligent transportation and automotive industry research communities. A digital parking map with precise parking spot geospatial information is crucial for research such as automatic valet parking, parking spot recommendations, and parking route optimization. Currently, these maps are generated through costly and resource-intensive human annotation. To address the challenges of insufficient parking maps and the high costs of manual map creation, researchers have proposed the following two strategies for generating parking maps. The SLAM-based map generation for indoor parking lots and the remote sensing image-based map generation for outdoor parking lot. Our work mainly focuses on the outdoor high-definition parking map generation from remote sensing images. These images often suffer from occlusion, inconsistent resolution, and varying luminosity conditions. To

Shukla, Ajitesh, Cao, Xiaofei, Liu, Yongkang, Takeuchi, Yusuke, Sisbot, Akin

A Real-Time Repositioning Strategy for SAEV Fleets Using a Simulation-Informed Deep Learning Model

2026-01-0035To be published on 04/07/2026

Shared Autonomous Electric Vehicles (SAEVs) can enhance urban mobility and efficiency. However, their operational performance is often hindered by the spatio-temporal imbalance between vehicle supply and passenger demand, leading to long wait times. This paper develops a novel repositioning framework where a lightweight CNN, informed by computationally intensive multi-agent simulations, enables real-time strategy deployment. The results show that: (1) An optimized repositioning policy, calibrated via multi-agent simulation, effectively cuts the mean passenger waiting time from 12.0 to 3.0 minutes (a 75% reduction). (2) A lightweight CNN surrogate model enables real-time deployment, reducing the policy computation time from ~4 hours to ~5 minutes (>98% faster). (3) The deep learning surrogate achieves this speed with a negligible performance trade-off, increasing the waiting time by only 0.156 minutes (4.9%) compared to the full optimization.

Shang, Kai, Wang, Ning

Autonomous Racing Control Using Reinforcement Learning with Racing Line Guidance

2026-01-0031To be published on 04/07/2026

In recent years, with the emergence of autonomous racing competitions such as Roborace, autonomous racing cars have become a prominent research focus and have undergone rapid development. This paper establishes a reinforcement learning-based decision-making and control framework for autonomous racing cars, achieving cooperative optimization between high-speed racing and safety control under extreme operating conditions. By integrating a Control Barrier Function constrained reward function for dynamic stability and a racing line-guided path optimization, the proposed framework achieves a synergistic balance between minimum lap time and operational safety. Dynamic stability domain constraints based on a two-degree-of-freedom single-mass rigid-body model and racing line optimization based on track geometric modeling are established. The phase plane stability boundary is constructed using the yaw rate method and double-line method, upon which Control Barrier Function constraints are

peng, tao, Zhang, Weiqi, Li, Zengwen, sudong, jiang

ProGuard: A Real-time Preemptive Artificial Intelligence-Based Anti-Pinch System for Bus Entryways on Low-Power Edge Devices

2026-01-0100To be published on 04/07/2026

This paper proposes ProGuard, a novel approach to preemptive pinch detection systems for buses. ProGuard utilizes state-of-the-art AI object detection algorithms to identify potential pinching events in bus entryways before pinching occurs. Modern conventional anti-pinch systems, such as pressure sensors or hall effect sensors, often rely on mechanical contact before triggering. While these systems are established safety mechanisms, they are reactive and therefore require some level of pinching before triggering. This reactive approach presents numerous safety concerns for passengers, especially when considering children on school buses. Existing preemptive detection methods, such as infrared or ultrasonic sensors, solve the problems presented by these reactive detection systems. However, these systems either lack the range or environmental resilience needed for reliable operation in buses. The critical nature of anti-pinch systems requires a robust and reliable solution that can adapt

Bradley, Hudson, Zadeh, Mehrdad, TAN, Teik-Khoon

Ford Vehicle Lines Comprehensive Sales Forecasting: A Machine Learning Approach with Price Sensitivity Analysis

2026-01-0108To be published on 04/07/2026

The automotive industry faces constant challenges in accurate demand prediction due to market volatility, macroeconomic fluctuations, and the complexity of consumer purchasing decisions. This project develops a comprehensive forecasting system for all Ford vehicle lines and their respective catalogs, implementing an advanced machine learning methodology that integrates endogenous and exogenous variables to generate 12-month sales predictions with price sensitivity analysis. The primary objective is to develop a robust predictive model capable of forecasting sold units for each Ford vehicle catalog under different price variation scenarios, providing strategic insights for commercial and pricing decision-making. The system enables evaluation of price increases or decreases impact on projected demand, facilitating pricing strategy optimization and revenue maximization. The model architecture incorporates a hybrid approach combining time series, neural networks, and advanced machine

Hernandez, Alejandro, Hernandez Cervantes, Oscar, Landa, Ruth

Open Source LLM Performance in Automating Embedded C Software Quality Improvements for Automotive

2026-01-0103To be published on 04/07/2026

The study presented in this paper explored the potential of five open-source Large Language Models (LLMs) with parameter counts between 30 billion and 50 billion to automate enhancements in code quality and developer productivity. The evaluated models - CodeLlama, Command-R, Deepseek R1, Nemotron, and QwQ - were assessed on their ability to refactor a large and complex automotive mechatronic C language function. This assessment focused on adherence to provided code quality standards and successful compilation of the refactored function within its original code module. The evaluation also compared the impact of parameter count, hyperparameter tuning, model architecture, and fine-tuning. This comparison revealed that larger models generally outperformed smaller models against the provided quality standards. Additionally, hyperparameter tuning yielded a moderate improvement in performance. The study further highlighted that model architecture and fine-tuning had less predictable effects

Struck, Daniel, Kumaraswamy, Samanth

Real-time Parameter Optimization for Eco-driving Control in Connected and Automated Vehicles Using Reinforcement Learning

2026-01-0041To be published on 04/07/2026

This paper introduces a novel methodology to enhance the energy efficiency of eco-driving controllers in Connected and Automated Vehicles (CAVs) by leveraging Reinforcement Learning (RL) techniques for real-time parameter optimization. Traditional eco-driving strategies rely on fixed control parameters, which limit adaptability across diverse traffic and road conditions. To address this, we apply continuous action space RL algorithms, specifically Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO), to dynamically tune four key parameters within a model predictive control framework that is grounded in Pontryagin’s Maximum Principle (PMP). These parameters influence acceleration, braking, cruising, and intersection-approach behaviors, making them critical for achieving optimal eco-driving performance. Our study employs Argonne National Laboratory’s RoadRunner simulator, a Simulink-based environment designed for high-fidelity CAV analysis, incorporating

Zhang, Yaozhong, Ammourah, Rami, Han, Jihun, Moawad, Ayman, Shen, Daliang, Karbowski, Dominik

Roundabout Dilemma Zone Data Mining and Forecasting with Trajectory Prediction and Graph Neural Networks

2026-01-0029To be published on 04/07/2026

Traffic roundabouts, as complex and safety-critical road scenarios, present significant challenges for autonomous vehicles. In particular, predicting and managing dilemma zone (DZ) encounters at roundabout intersections remains a pivotal concern. This paper introduces an AI-driven system that leverages advanced trajectory forecasting to anticipate DZ events, specifically within traffic roundabouts. At the core of our framework is a modular, graph-structured recurrent architecture powered by graph neural networks (GNNs). By modeling agent interactions as a dynamic graph, our approach integrates heterogeneous data sources, including semantic maps, while capturing agent dynamics with high fidelity. This GNN-based forecasting model enables accurate prediction of DZ events and supports safer, data-driven traffic management decisions for both autonomous and human-driven vehicles. We validate our system on a real-world dataset of roundabout intersections, where it achieves high precision with

Lu, Duo, Farhadi, Mohammad, Satish, Manthan, Yang, Yezhou, Chakravarthi, Bharatesh

Displacement measurement in impact tests based on machine vision

2026-01-0477To be published on 04/07/2026

Contact measurement is difficult to arrange sensors under various impact tests. And the method of measuring displacement by machine vision has significant advantages. This paper proposes a displacement measurement method based on machine vision. Firstly, YOLO (You Only Look Once) v5 is used to automatically extract the recognition targets in various test scenarios. Secondly, the extracted recognition target graphics are processed by image enhancement and threshold segmentation technology. Then, the geometric shape of the processed graphics is identified by edge detection technology, and the scale factor is calculated. Finally, the displacement of the target structure is calculated through coordinate transformation in multiple coordinate systems. The results show that for a single recognition target, the relative error between machine vision and sensor measurement is within 5%, and for multiple recognition targets, the relative error between machine vision and industrial software

Sun, Jiawei, Du, Dan, Zhang, Bin

Generative Agents for High-Fidelity Simulation of Community-Scale Mobility and Energy Behavior

2026-01-0465To be published on 04/07/2026

The transition to sustainable mobility and energy systems is a complex socio-technical challenge, where the success of new technologies and policies hinges on their interaction with human behavior. Traditional models often fail to capture the nuanced, heterogeneous, and adaptive nature of occupant decision-making, particularly in terms of mobility choices and energy use, creating a critical uncertainty gap for system design and policy evaluation. This paper presents a new paradigm to address this gap: a scalable, high-fidelity simulation framework based on Generative Agents. The framework's core innovations are twofold: 1) It leverages Large Language Models (LLMs) to autonomously generate agents with intrinsic personality traits autonomously, enabling a realistic simulation of diverse, human-like responses to environmental, social, and transportation-related cues. 2) It introduces a hierarchical "Prototype-Individual" architecture that makes large-scale simulation of entire communities

Chen, Yongjian, Yang, Zhifeng, Ou, Shiqi(Shawn)

A Large Language Model-Based Database for Analyzing the Electric Vehicle Battery Supply Chain

2026-01-0471To be published on 04/07/2026

Global geopolitical volatility poses a critical threat to the resilience of the electric vehicle battery supply chain. Static, manually updated databases fail to capture the sector's rapid dynamics, creating significant information gaps for strategic planning. This study proposes an Artificial Intelligence-driven methodology to construct a comprehensive and dynamic knowledge base. An automated pipeline is implemented, beginning with specialized web crawlers that harvest real-time textual data from curated news and industry sources. This unstructured data is preprocessed through deduplication and cleansing to ensure quality. The core innovation involves applying fine-tuned Large Language Models (LLMs) to perform deep semantic parsing and extract structured information. These models precisely identify key entities—corporations, facilities, production capacities—and delineate the complex multi-tier relationships from raw material extraction to final distribution. The output is an

Zhu, Juntong, Luo, Wei, Zhang, Xiang, Yang, Zhifeng, Ou, Shiqi(Shawn), He, Xin

Optimization of a Feedforward Neural Network Machine Learning Battery Model for an Automotive Pouch Cell

2026-01-0386To be published on 04/07/2026

Modern battery management systems have a critical need for highly accurate battery terminal voltage models, which are a key component of algorithms which estimate or predict power capability, range, temperature, and other factors. While electro-chemical and equivalent circuit models are widely used for this purpose, they typically struggle to capture the complex, non-linear dynamics and degradation effects inherent in real-world battery operation, necessitating frequent recalibration. This study addresses these limitations by proposing a robust, data-driven approach for terminal voltage estimation using a feedforward neural network (FNN) machine learning model. An LG E66 cell, a high-performance lithium-ion battery with a Nickel Cobalt Manganese (NCM) cathode, was selected for this study due to its prevalent use in light duty electric vehicles. Data collected at temperatures ranging from 40 °C to -20 °C is used to train and test the models, with model error reported for HWFET, UDDS

Dehury, Biswanath, Nahidmobarakeh, Lucas, Panchal, Satyam, Gross, Oliver, Kollmeyer, Phillip

Machine Learning Modeling for Compressive Property Prediction of Foams

2026-01-0512To be published on 04/07/2026

Foam material models for automotive structural analysis typically require tensile and compressive data at multiple strain rates. The testing is costly and may require a long time to complete. For many applications, foams of similar chemistry are used and the foam structural responses, such as stiffness and compression force deflection, are controlled by the foam density. In such cases, Machine Learning (ML) lends itself as an ideal tool to detect the trends in material response based on density and strain rate. In this paper, five microcellular polyurethane (PU) foams of different densities were tested at four strain rates ranging from 0.01/s to 100/s. A ML model capable of predicting compressive stress-strain response for a range of densities was developed. The model demonstrated good prediction capability for intermediate strain rates at all foam densities and in extrapolating stress-strain curves at higher densities at all strain rates. The strain rate trends for a density outside

M, Gokula Krishnan, Kavimani, Harish, Muppana, Sai Siddhartha, Savic, Vesna, Chavare, Sudeep, V S, Rajamanickam

Design and Evaluation of a Deep Neural Network Model for Prediction of Diesel Equilibrium Combustion Products and Characteristics

2026-01-0293To be published on 04/07/2026

Accurate prediction of equilibrium combustion products and thermodynamic properties is essential for optimizing engine performance, enhancing combustion efficiency, and reducing emissions in diesel-powered systems. Traditional methods for combustion modeling often involve solving complex chemical equilibrium equations or thermodynamic relations, which could be computationally expensive and time-consuming. In this study, we present a data-driven approach using a deep neural network (DNN) model to predict the equilibrium combustion products and key thermodynamic characteristics of diesel under varying thermodynamic conditions. The proposed DNN model is trained on a comprehensive dataset generated from equilibrium calculations. The inputs include pressure, temperature, and equivalence ratio, covering a relatively wide range to encompass diesel equilibrium combustion under various conditions. Outputs are equilibrium combustion products and thermodynamic properties, including enthalpy

Ji, Huangchang, Wang, Kai, Guo, Zhefeng, Han, Yang, Lee, Timothy

Items per page:

1 – 50 of 2399