Browse Topic: Neural networks

Items (1,486)
In this study, an efficient method for concurrent thermomechanical performance and weight optimization under modal constraints is proposed to address the coupled design challenges of thermomechanical characteristics (thermal capacity, thermal deformation, and modal) and structural weight in straight-ribbed brake discs. Based on high-fidelity computer-aided engineering (CAE) simulations of brake disc thermomechanical behavior, a neural network (NN)-based surrogate model and a ResNet-guided geometric feature recognition (RGFG) model for automatic modality recognition were developed, and integrated with a particle swarm optimization (PSO) framework for optimal solution exploration. When applied to a passenger vehicle brake disc case study, the surrogate model of NN demonstrates remarkable accuracy: it shows more than 95% agreement with the CAE results in thermal capacity prediction, the prediction accuracy of thermal deformation exceeds 90% compared to CAE results and 83.4% compared to test result, thereby validating the method’s effectiveness. Compared with conventional CAE approaches, the surrogate model of NN achieves a subsecond prediction speed, significantly reducing computational costs. The surrogate model of RGFG achieves a test accuracy exceeding 95%. Furthermore, the proposed optimization framework offers valuable insights for the inverse design of brake discs.
Han, SimiaoJiang, DaxinHan, ChaoWang, JindaSui, Qinghai
Terminal guidance is critical for ensuring strike precision in the final phase of flight. However, traditional methods, such as proportional navigation and optimal guidance laws, face significant challenges regarding real-time performance and adaptability to dynamic targets. To address these issues, neural networks offer a promising solution by enabling adaptive adjustments to guidance parameters, thereby improving performance under various constraints.
Ma, HengweiWang, YongfengWen, HongLiu, DiWei, YuanhangDong, LonghaoLuo, Ying
In map-free geomagnetic navigation conditions, the traditional matching algorithms will be ineffective, and the regular position searching optimization algorithms still face the problems of low navigation accuracy and inefficiency. How to further improve the accuracy and efficiency of the algorithm has become the key to the application of this method in maple’s geomagnetic navigation conditions. Based on the above background, this paper proposes an evolutionary gradient search navigation algorithm optimized via position estimation (PE-EGA). The world geomagnetic model (WMM) is used to establish the nonlinear correlation relationship between geographic position and geomagnetic features, and the inverse mapping of the geomagnetic model is fitted by a fully connected neural network to get the rough estimation of the geographic position of the vehicle, with a root mean square error (RMSE) of 0.0121 in position estimation. Finally, the information of the rough estimation is used to assist the decision-making of the navigational azimuth angle involved in the EGA algorithm. The simulation results show that the offset distance of the improved algorithm is only 27.09 m, and the path ratio reaches 1.0178 with an error ratio of 0.38%. Comparative study using measured geomagnetic data of Boao town with model data shows that the final offset distance is only 51.63 m, path ratio 1.0036, and error ratio 0.73%, which significantly improves the accuracy and timeliness of navigation compared to the original EGA algorithm. This article provides an innovative and practical solution strategy for map-free geomagnetic navigation.
Xie, WenbinLiu, HongjieZheng, RuifanRen, XintianYan, BingQiu, WeiChen, Zhuo
This paper constructs a reinforcement learning framework based on the PPO algorithm for drone air combat to solve 1v1 pursuit-evasion in 2D beyond-visual-range air combat. Firstly, the mission scenario is modeled, defining key roles of ATA and AA. Then, state transition models of pursuer and evader are built based on flight kinematics. To handle reward sparsity in policy network training, a dense reward function combining distance and angle rewards is designed to guide the agent in learning tail-chasing and interception strategies. Using the Actor-Critic architecture, deep neural networks implement the decision-making and evaluation modules. The PPO algorithm trains the pursuing drone in a simulation. Results show that after ~5 million steps, the agent learns a stable strategy, completing tasks promptly and generalizing well in unseen scenarios. This research offers ideas for drone combat and guidance, and supports autonomous decision-making in complex air battles.
Yu, KangjieGong, ZhengHu, RunchangLiu, Huixiang
Efficient optimization of aerodynamic shapes is a critical challenge in aircraft design. Traditional CFD-based optimization workflows suffer from high computational costs and low efficiency, which severely restricts their practical engineering application. In this paper, a novel aerodynamic optimization method based on a hierarchical neural network with adaptive activation functions is proposed. The network adopts learnable B-spline activation functions and is hierarchically constructed in accordance with the sharing status of B-spline control points. After being trained to achieve fast and accurate prediction of aerodynamic performance, the network can effectively replace the traditional CFD module in the optimization loop. The primary advantage of the proposed method is that it significantly reduces the computational cost during the optimization process while ensuring that the prediction accuracy is not compromised. This work thereby presents a novel strategy and technical framework for streamlining the design process of hypersonic vehicles.
Liu, DiWang, YongfengWen, HongWei, YuanhangMa, HengweiZhao, Runhui
This paper presents an innovative study in exploring, evaluating, and implementing deep-learning architectures for the calibration of multimodal sensor systems. The aim of this paper is to leverage the use of sensor fusion to achieve dynamic, real-time alignment between 3D LiDAR and 2D camera sensors. Static calibration methods are tedious and time-consuming, which is why we propose utilizing conventional neural networks (CNNs) coupled with geometrically informed learning to solve this issue. We leverage the foundational principles of extrinsic LiDAR–camera calibration tools such as RegNet, CalibNet, and LCCNet by exploring open-source models that are available online and compare our results with their corresponding research papers. Requirements for extracting these visual and measurable outputs involved tweaking source code, fine-tuning, training, validation, and testing of each of these frameworks for equal comparisons. This approach aims to investigate which of these advanced networks produces the most accurate and consistent predictions. Through a series of experiments, we reveal some of their shortcomings and areas for potential improvements. We find that LCCNet yields the best results among all the models that we validated.
Karramreddy, Venkat Sai RaxitMitchell, Liam
This paper puts forward a Privacy-Preserving UAV-Based Traffic Data Acquisition Platform to address 1) privacy leakage, 2) limited scenario coverage, and 3) low traffic data utilization efficiency in urban traffic monitoring environments. Our system integrates three innovations: 1) Dynamic Privacy Masking (DPM) and Dual-Track acquisition (DTC), which hides sensitive information (e.g., faces, license plates or LPL) in real-time while preserving critical traffic data (e.g., vehicle density, speed), 2) traffic data Localization (DL) and Privacy-Enhanced Federated Learning (FEFL), enabling cross-regional collaboration without raw traffic data sharing by perturbing neural network updates with differential privacy (DP), and 3) Ground-Air Collaboration (GAC) and VPF (VPF), combining UAVs with ground sensors and digital twins (DTs) to cover blind spots (e.g., tunnels, extreme weather). Experimented on UA-DETRAC and CitySim traffic data-sets, the platform achieves 92% privacy compliance (GDPR/PIPL), 87.5% mAP accuracy, and 85% road network coverage, outperforming other methods (e.g., FedUAV, static blurring). It supports applications such as traffic flow optimization, accident prevention, and regulatory alignment.
Zhang, ShilinYan, Ming
With the rapid development of the low-altitude economy—represented by drone logistics, aerial inspections, and air taxis—air traffic has exhibited new characteristics including diverse forms, high density, and significant speed differences. To address these changes, the traditional air traffic control system requires upgrades, particularly in dynamic aircraft scheduling. This study proposes an air traffic control model (DS-ATM) tailored to this domain, built on the Deepseek large model. By integrating spatiotemporal graph neural networks with multi-objective reinforcement learning algorithms, the model achieves real-time path planning and conflict resolution in complex airspace environments. Validated using public datasets such as OpenSky Network, NASA UTM Dataset, and METAR meteorological data, experimental results demonstrate its significant advantages in reducing conflict rates and scheduling delays.
Li, RuiZhao, FangyuShe, YueLi, Wujie
This paper reviews data fusion strategies for generating aerodynamic databases and evaluates their suitability for motorsport aeromaps, with emphasis on the operational constraints specific to Formula One. A structured survey and classification of the state of the art is presented, grouping approaches into (i) surrogate-agnostic methods, (ii) kriging-based methods, and (iii) neural network–based methods. In addition, the survey explores advanced techniques currently underutilized in aerodynamic database applications but that show promise. These methodologies are discussed in the context of addressing limitations inherent in traditional approaches, such as dependency on nested sampling plans and linear correlation assumptions between low- and high-fidelity datasets. The review indicates that, although multi-fidelity data fusion is well established in aerospace aerodynamic database generation, its direct translation to motorsport requires additional considerations. In the Formula One context, the most plausible deployment may involve fusing legacy and current datasets, rather than combining low- and high-fidelity evaluations of the same geometry. This shift in premise could increase exposure to negative transfer and therefore necessitate additional methods to minimize it. This study provides one of the first motorsport-focused reviews and syntheses of data fusion methods for aerodynamic database generation. It is intended to guide motorsport engineers and researchers toward more efficient and effective aeromap generation strategies. Collectively, the findings establish a foundation for subsequent phases of a broader project to minimize the number of data points required to generate an aeromap, with the present survey constituting the first part of that effort.
Ongley, Thomas James HenryTeschner, Tom-RobinAshton, NeilSiampis, Efstathios
Software-defined, highly customizable vehicle architectures drastically increase the number of hardware–software constellations that must be validated, especially under safety and timing constraints. Traditional unit and integration testing, as well as current regression and combinatorial methods, cannot practically cover this configuration space or reliably capture emergent effects arising from complex interactions, such as bandwidth contention and non-linear latency behavior. This work presents a proof-of-concept for predictive, situational validation of self-describing hardware and software components within realistic automotive E/E architectures. Proposing a novel Machine Learning- (ML) based method for early systemic feasibility prediction of automotive configurations using Graph Neural Networks (GNNs). Specifically, the subclass Graph Isomorphism Networks (GINs) is applied to predict the compatibility of a randomly composed configuration of software and hardware components, assessing both structural compatibility and functional stability. The trained models achieve recall and accuracy above 90%, even when detailed behavioral metadata is hidden during training, indicating that systemic incompatibilities are learnable from topological features alone. Results were achieved from training on a realistic, synthetic dataset representing less than 10e−27% of all possible permutations without finetuning or further parameter optimization. It demonstrates the potential of GIN-based graph learning to enable early, automated feasibility assessment, substantially reducing testing time and development effort for modular, personalized, and update-capable vehicle architectures.
Wizl, JensGuarda, Filippo
Electronic Control Units (ECUs) have played a pivotal role in transforming motorcars of yore into the modern vehicles we see on our roads today. They actively regulate the actuation of individual components and thus determine the characteristics of the whole system. In this, the behavior of the control functions heavily depends on their calibration parameters which engineers traditionally design by hand. This is taking place in an environment of rising customer expectations and steadily shorter product development cycles. At the same time, legislative requirements are increasing while emission standards are getting stricter. Considering the number of vehicle variants on top of all that, the conventional method is losing its practical and financial viability. Prior work has already demonstrated that optimal control functions can be automatically developed with reinforcement learning (RL); since the resulting functions are represented by artificial neural networks, they lack explainability, a circumstance which renders them challenging to employ in production vehicles. In this article, we present an explainable approach to automating the calibration process using residual RL which follows established automotive development principles. Its applicability is demonstrated by means of a map-based air path controller in a series control unit using a hardware-in-the-loop (HiL) platform. Starting with a sub-optimal map, the proposed methodology quickly converges to a calibration which closely resembles the reference in the series ECU. The results prove that the approach is suitable for the industry where it leads to better calibrations in significantly less time and requires virtually no human intervention.
Kampmeier, AndreasBadalian, KevinKoch, LucasLee, Sung-YongAndert, Jakob
The optimization of energy management strategies for hybrid electric vehicles is crucial for minimizing fuel and electrical energy consumption while maintaining the energetic stability of the electrical system. Conventional heuristic, rule-based approaches typically rely on classical optimization techniques and manual calibration by experienced engineers. These methods often suffer from simplified assumptions, sub-optimality, and are increasingly time-consuming given the growing complexity of modern hybrid powertrain architectures. This research proposes a novel methodology for the development of a learning-based energy management strategy (EMS) via deep reinforcement learning (DRL) to transition toward highly automated, data-based, and optimization-based development approaches. The methodology utilizes the Soft Actor-Critic (SAC) algorithm, an off-policy actor-critic method, to train an agent through experiences by interacting with an environment. The environment consists of a backward-looking, quasi-static vehicle longitudinal dynamics simulation model of an exemplary P2 plug-in hybrid electric vehicle (PHEV) combined with a database of customer-representative driving profiles. The agent learns optimal control policies through defined states and actions, optimizing a multi-criteria reward function that balances fuel efficiency against energetic stability. The framework permits the definition of both non-predictive and predictive states. Additionally, a shield function is implemented to consider hard constraints ensuring safe and stable operation. Variation calculations and sensitivity analyses regarding reward function shaping and hyperparameter tuning are conducted. The agent is trained in an offline simulation environment, and the learned policy of the trained deep neural network (DNN) is transferred into deterministic control maps, applicable to vehicle control units, ensuring interpretability, reproducibility, and compliance with certification requirements. Finally, exemplary simulation results of the DRL-EMS approach are presented and compared to benchmark equivalent consumption minimization strategy (ECMS). In conclusion, the proposed methodology enables a generally valid approach for the development of learning-based energy management strategies towards close-to-optimal strategies while reducing manual calibration effort.
Metzler, SebastianWinke, FlorianJungen, MarioSchmiedler, StefanHofmann, PeterGeringer, Bernhard
Kolmogorov-Arnold Networks (KANs) are a novel mathematical method to generate data-driven AI surrogate models. Compared to neural networks based on the MLP standard (Multi-Layer Perceptron), these offer further mathematical interpretability and thus allow improved validation of AI for industrial applications. In this paper, we use KANs to generate an AI vehicle model of a truck as a mathematically precise AI surrogate model. To do this, we combine the KAN approach with the approach of Neural Ordinary Differential Equations (Neural ODEs) to generate predictions for the time-series of the truck’s velocity. Furthermore, we compare the results of the AI based on KANs with the traditional approach using MLP in terms of model size, accuracy, and computational time in order to evaluate advantages and disadvantages of the KAN approach. The best AI-KAN vehicle model identified in this way is then embedded in a co-simulation via the Functional Mockup Interface standard, thus opening up a wide range of applications in AI-driven truck development.
Vaudrevange, Patrick K.S.Halverson, JamesRuehle, FabianFabcic, TomazDingler, ChristianPiskala Dilipkumar, SanthoshkumarIbrahim, MuhammedHerrnberger, MichaelKasper, JohannaTürk, LarsKeckeisen, Michael
Achieving best-in-class Noise, Vibration, and Harshness (NVH) in electric powertrains demands a paradigm shift in development methodology. This paper presents a practice-oriented overview of simulation methods in NVH development methodology for electric drive units. This includes target cascading and multi-objective optimisation, and by attacking NVH at the source using KPIs early in the design cycle, significant reductions in development time and reliance on traditional testbed loops are realised. Machine learning (Neural Network) algorithms are utilized to find the best-in-class design, using multi-objective optimisation as well as refining simulation accuracy by adding tolerance effects while target cascading ensures alignment of system-level performance objectives down to subsystem contributions. Combined, these strategies enable rapid and robust NVH optimisation, using simulation for next-generation electric powertrain development. Several applications and real-life examples demonstrate how simulation helped with NVH issue identification or improvement.
Mehrgou, MehdiGarcia de Madinabeitia, InigoGraf, BernhardGojo, Josef
In this study, we propose a methodology for predicting the acoustic modes and natural frequencies of a sedan using artificial intelligence and demonstrate the feasibility of controlling its acoustic characteristics by modifying the hole distribution of the package tray. In typical sedan structures, the cabin cavity and trunk cavity are acoustically coupled through holes in the package tray. The distribution of these holes significantly affects the natural acoustic modes and frequencies of the vehicle. However, once the exterior shape of the vehicle is finalized during the design stage, options for structural modifications to mitigate noise issues caused by these modes become extremely limited. To address this challenge efficiently, we develop a deep learning-based neural network model trained on data derived from a simplified acoustic analysis model of a sedan that includes a package tray. Finite element analysis is performed to generate acoustic modes and natural frequencies, which serve as training data, for various hole distributions. The trained model is then used to predict acoustic natural modes and natural frequencies from unseen input images representing different hole configurations in the package tray. These predictions are made in a fraction of the time required for traditional simulation methods, thereby validating the model’s effectiveness. Furthermore, we demonstrate that the latent variables embedded in the trained model can be manipulated to control the acoustic modes and natural frequencies of the sedan. This indicates the potential for artificial intelligence-driven acoustic design optimization in early-stage vehicle development, offering both time efficiency and design flexibility without physical prototyping or extensive simulations.
Lee, Jin WooCho, JaehoNam, YounsicHan, Yongha
In recent years, the automotive industry has actively explored the application of various AI-based models such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, Autoencoders, and Transformers to improve defect detection rates at the End-of-Line (EOL) stage. However, implementing these approaches in the Noise, Vibration, and Harshness (NVH) area face several practical challenges: ① extended evaluation times compared to other data types, which limit the quantity of training data and lead to overfitting; ② label imbalance caused by the relatively small amount of defect data; ③ reduced labeling accuracy due to human error; ④ decreased robustness under domain shifts such as changes in jig fixtures, test environments, and signal-to-noise ratio (SNR); ⑤ diminished model reliability when new defect arise during development; and ⑥ constraints imposed by compatibility requirements with existing test equipment. This study proposes a Convolutional Autoencoder (CAE) based framework trained on NVH datasets collected from normal and defective Column-type Electric Power Steering (C-EPS) systems. Latent variables at the bottleneck layer are used for dimension reduction, enabling visualization and unsupervised classification using a clustering algorithm. A classification model derived from the encoder is fine-tuned with clustered data, and Gradient-weighted Class Activation Mapping (Grad-CAM), an eXplainable AI (XAI) technique, is applied to extract Feature Frequency Maps (FFM) highlighting defect-related noise and vibration characteristics. The proposed approach does not rely on the deep learning model to directly classify defect. Instead, it utilizes extracted FFM as weights(mask) to detect defect. This method enables quantitative data representation and ensures high applicability with existing EOL equipment. Post-processing within the FFM enables root cause analysis, reducing issue resolution time and supporting integration with conventional signal analysis techniques.
Park, Jun-SeoJo, Hyeon-ChoelCho, In-JeSeo, Jae-YongYoo, Seong-Sik
Neural Network Enabled Synthetic Air Data System: Development and Validation2026-26-07206/1/2026
Synthetic Air Data System (SADS) provides a smart solution that can be used to predict critical air data parameters in the absence of conventional air data sensors. Traditional air data sensors, such as pitot-static tubes and vanes, are generally expensive, require regular maintenance, and can fail in harsh weather conditions. In addition, these sensors, along with their processor and computers, add weight to the aircraft. To address these issues, a synthetic air data system is proposed using a Recurrent Neural Network (RNN). Several flight variables were checked for Pearson correlation coefficient with respect to the angle-of-attack and angle-of-sideslip, and thereafter, input features were selected based on the thresholding technique. The proposed neural network has two hidden layers and regularization technique was implemented by adding two dropout layers to each hidden layer to prevent overfitting of the model. The neural network was trained using actual flight test data, supplemented with simulated data wherever gaps were observed in the entire flight envelope. The RNN model is trained to predict the aerodynamic flow angles, viz., angle-of-attack and angle-of-sideslip. The proposed model was found to be able to predict the aerodynamic angles with a degree of accuracy. The accuracy was also checked with several complementary actual flight data to check the fidelity of the trained neural network model.
Sahu, SanjuC, PoornimaKaliyari, DushyantTK, Khadeeja NusrathHebbar, Archana
The electrical harness system of satellite launch vehicles functions as the backbone of spacecraft avionics; inter connecting subsystems through complex networks of wires and connectors. An electrical harness is a group of wires bunched together and terminated in connectors. The common insulations used for launch vehicle applications include PTFE, Polyimide, ETFE and TKT. The connectors used are of aerospace grade and connectors tailored for space applications. With over 5000 connectors and 200 km of cables constituting nearly 20% of vehicle mass, the design, fabrication, and sustainability of these systems are critical. The insulations of connectors inserts or the wires are critical for the durability of harness elements. Nevertheless, these insulations are non-expendable and pose disposal challenges and some releases toxic gases when burned or due to vacuum outgassing phenomenon. Also, the cadmium plating which is often used for the environmental resistance of connector shells presents additional risks to the working humans due to its carcinogenic nature and shows tendency to bloom out during storage. This paper presents the methodologies and innovations implemented to develop safe, reliable, and environmentally sustainable harness systems for current and future launch vehicles. Key advancements include the adoption of lean manufacturing practices for waste reduction, the replacement of hazardous cadmium-plated connector shells with stainless-steel alternatives, and the induction of TKT-insulated wires to prevent arc tracking and ensure human-rating compatibility. Additionally, lightweight composite connectors and micro-miniature interconnects are being qualified to support mass optimization in reusable launch vehicles. Through these strategic measures, the study demonstrates how the integration of sustainable materials, safety-oriented design, and process optimization can enhance the performance, safety, and environmental footprint of launch vehicle electrical harness systems.
K S, NithishTR, BinnyD S, Praveen Kumar
Unscheduled maintenance due to the failure of critical components, such as aero-engine rolling element bearings, is a leading cause of costly Aircraft-on-Ground (AOG) events; consequently, current time-based maintenance practices are inefficient and prone to risk. This paper develops a resource-efficient Hybrid Digital Twin (HDT) model for an engine bearing, focusing on the dynamic prediction of spall growth due to Rolling Contact Fatigue (RCF), thereby enabling a condition-based maintenance paradigm. The HDT architecture integrates two core models: (1) a physics-informed model that uses established life and fatigue theory to define initial degradation thresholds, and (2) a data-driven Recurrent Neural Network (RNN), specifically a Long Short-Term Memory (LSTM) network, for dynamic degradation rate modeling. The methodology utilizes a Monte Carlo simulation coupled with RCF progression equations to generate a large, high-fidelity synthetic run-to-failure dataset under varying operational loads, accurately simulating realistic mission profiles. This approach addresses the critical "data scarcity" challenge in aviation. To ensure operational reliability, the framework incorporates Uncertainty Quantification (UQ) using Monte Carlo Dropout and addresses the "Sim-to-Real" gap through Transfer Learning on the NASA IMS bearing dataset. The HDT demonstrates a significant improvement in prognostic accuracy, achieving a Root Mean Square Error (RMSE) reduction of over 71% compared to baseline models. Furthermore, a cost-benefit analysis suggests a potential fleet savings of $240,000 per 100 engines by avoiding false negatives. This computationally efficient approach supports the Digital Engineering Transformation theme by providing a scalable blueprint for the virtual qualification of critical mechanical components.
Mohamed, Abbas
This paper addresses the critical challenge of fault-tolerant control in autonomous multi-copters, particularly under conditions of one or two rotor failures a scenario that often leads to severe instability and a complete loss of directional control due to unbalanced torque and resultant autorotation. Existing advanced control strategies, including optimal approaches such as LQR, typically require precise system modeling and state estimation, which are difficult to achieve in real-world, dynamic failure scenarios. Alternative methods like fuzzy logic, sliding mode control, and gain-scheduling either lack robust generalization or are impractical for enumerating all possible failure cases. In this work, a hybrid control framework integrating Physics Informed Neural Networks (PINN) with a standard PID controller is proposed for fault-tolerant operation of autonomous multi-copters subject to multiple actuator failures. PINNs incorporate governing physical laws as regularization in their loss functions, allowing them to learn optimal counter-torque actions and thrust balancing necessary to arrest autorotation and stabilize flight, despite limited training data and uncertainty in failure conditions. The calculated moments and thrust commands are executed via a robust PID scheme, enabling reliable real-time implementation and minimizing residual oscillations. This hybrid control architecture demonstrates significant potential to enhance the resilience and operational safety of autonomous multi-copters during unexpected motor failures. By leveraging PINN’s physics-based generalization and PID’s consistent execution, the proposed method offers an adaptive, model-agnostic approach for maintaining stable flight and directional control under severe actuator faults, with implications for next-generation fault-tolerant UAV systems deployed in complex environments.
Charapalle, SamruddhiVenugopalan, NandagopalanNerkundram Muralidharan, ArunSundararaj, Laveen
With the introduction of China’s dual-carbon goals (carbon peak and carbon neutrality), renewable energy has experienced rapid development in the country, particularly wind energy, which has established a pivotal role within the new energy sector. However, the inherent fluctuations in wind power generation pose significant challenges to maintaining grid stability and operational reliability. In power systems where the proportion of installed wind power capacity has significantly increased, the allocation of flexible resources becomes crucial. These resources help the system adapt to fluctuations in wind power generation and load demand, avoid wind power curtailment, and reduce costs. In addition, energy storage enhances grid flexibility and stabilizes renewable energy, but is constrained by high costs. Therefore, optimizing energy storage allocation and improving its economic efficiency have become urgent issues. This study focuses on flexibility adequacy assessment and resource allocation, proposing an optimization model that incorporates constraints from wind power, thermal power, energy storage, and load shedding. The model aims to identify the optimal energy storage configuration strategy to minimize operational costs while ensuring system flexibility adequacy. The model is constructed and simulated using MATLAB, with complex optimization problems solved using the Gurobi solver. The research involves developing flexibility evaluation indicators, determining the optimal energy storage capacity configuration, and creating an economic objective function. The goal is to minimize costs while ensuring system flexibility.
Peng, JianWei, JinpengZhu, ZhengyinHu, JianminLi, YuxiangMiao, GangZhang, Huaide
In order to achieve the research objective of simultaneously improving the air volume and reducing the noise of centrifugal fans, a combination of orthogonal experimental design, BP neural network modelling and multi-objective genetic algorithm (NSGA- II) was used to find the optimal method, and the worm tongue placement angle φ, worm tongue radius R, expansion angle θ and outlet expansion section height L of the worm casing were selected as optimization variables. The air volume and noise of the centrifugal fan under the design working condition were calculated by non-constant and constant calculations, and the air volume and noise were used as the optimization objectives. The results demonstrate that, compared to the initial design, the optimized fan model achieved a noise reduction of 10.99 dB and an airflow increase of 1.76%. Furthermore, the amplitude of the pressure pulsation coefficient at the blade passing frequency was significantly reduced at the monitoring point near the volute tongue. This suggests a reduction in the intensity of dynamic and static interactions between the impeller and the volute tongue, thereby enhancing the operational stability of the fan. The proposed optimization method has certain reference significance for improving the aerodynamic performance of centrifugal fans.
Huang, GuoxingZhang, WeihongLi, Weichang
Evaluating rotor component clearances is a multidisciplinary process aimed at ensuring that no contact occurs between rotor parts during a rotorcraft's operational life. It begins with calculating relative distances between components across all possible displacements and deformations combinations using a rotor kinematic model, and ends with clearance verification through flight data analysis and simulation. This task requires coupling detailed rotor aeroelasticity with flight mechanics to predict deformation under load, which is computationally expensive and unsuitable for real-time use. This work proposes a machine learning–based alternative: a neural network to estimate rotor clearances from flight mechanics inputs, with a specific application demonstrated in a simulated tiltrotor emergency maneuver with a pilot in the loop. The trained model successfully captures nonlinear relationships between maneuver parameters and rotor structural response, providing accurate predictions with reduced computational cost. The proposed framework can support both preliminary and detailed rotorcraft design and offers potential for real-time load prediction in applications where conventional aeroelastic tools are computationally expensive.
Zaccaria, AlessioOrsenigo, SimoneGerosa, GiacomoBergamasco, Marco
The proliferation of Autonomous Aerial Vehicles (AAVs) necessitates robust solutions for dynamic obstacle avoidance, particularly against non-cooperative intruders whose trajectories are unpredictable. While traditional path-planning algorithms excel in static environments, they struggle with dynamic obstacles due to the inherent difficulty in accurately estimating and registering their real-time depth and velocity into a world model. This paper presents a novel two-stage vision-based framework that leverages deep learning for reactive avoidance of non-cooperative dynamic intruders. Our approach decouples the perception and decision-making processes: an object detection deep neural network first processes monocular camera images to detect and track the 2D pixel coordinates of intruders. This perceptual output is then fed into a deep reinforcement learning agent, which learns a mapping from the intruder's image-space location to a high-level avoidance maneuver. This leads to more efficient learning, as the RL agent focuses solely on the policy without the burden of learning visual features. The advantage of using RL lies in its ability to handle partially observable situations—because reliable depth or full 3-D position information is not always readily available from monocular imagery, the RL agent learns to act based on the observable visual cues. Simulation results confirm that our proposed framework provides an effective solution for vision-based, non-cooperative intruder avoidance.
Dadkhah Tehrani, NavidWeintraub, JustinAmonkar, RikhilCarlson, SeanCherepinsky, Igor
This paper presents a spatio-temporal graph neural network (STGNN) centric approach to enable heterogeneous agents to collaborate and cooperate for different types of missions. The STGNN-centric approach and corresponding autonomy are encapsulated in the Advanced Graph-enabled Network Technology for Collaborative Autonomous Agents (AGENTCA) technology. Various decentralized and distributed control architectures are reported in the literature, but in some instances these approaches do not leverage the inherent graph network which can increase scalability to larger teams and algorithmic efficiency. Specifically, in this paper advances in artificial intelligence are leveraged to parameterize and encode optimal, or nearly optimal, swarm control techniques. For this work, the team focused on developing a diffusion-based STGNN swarm controller using imitation learning. An expert, centralized swarm control law was used to guide the STGNN during the learning process. The STGNN controller enables the swarm to follow a leader while avoiding static and dynamic obstacles and maintaining a desired separation distance from neighbors and obstacles. The approach is demonstrated in simulation with hundreds of agents and in flight tests with up to thirteen test vehicles.
Cooper, JaredLu, Chang-TienChen, SijiCarson, AndrewPeters, AndrewOlowin, AaronEnnasr, OsamaLichter, Matthew
This paper introduces a robust supervised machine learning framework for estimating helicopter gross weight during the takeoff phase. The methodology leverages high-fidelity datasets from Airbus's global in-service fleet to ensure a reliable training foundation. At the core of the approach is a long short-term memory recurrent neural network, supported by a patented data-curation pipeline designed to maintain high data integrity. To align with rigorous aviation safety standards, the study outlines a learning assurance process compliant with EASA guidelines, specifically addressing safety assessment objectives for machine learning. A central innovation is the characterization and monitoring of the model's operational design domain through multidimensional functional principal component analysis. By projecting high-dimensional, non-linear sensor data into a manageable tabular subspace, this approach enables the definition of safety envelopes using explainable and efficient classical methods. Validated against diverse real-world flight profiles, the framework demonstrates high predictive accuracy, marking a significant milestone toward deploying the model on airborne targets for safety-critical functions such as condition-based maintenance.
Mechouche, AmmarFabre, LouisValot, Nicolas
This paper focuses on the implementation of a novel supervised Machine Learning model for estimating helicopter weight during takeoff, utilizing extensive datasets from Airbus's global in-service fleet. The study details a learning assurance process aligned with the EASA concept paper for machine learning application, and with the on-going Eurocae ED-324. We propose a set of Machine Learning Requirements, a Machine Learning Model Description, and its implementation for a long short-term memory recurrent neural network. Finally, we verify the requirements on the implementation. Demonstrated on legacy avionics computers, the implementation is suitable for the deployment of the developed Machine Learning Model weight estimator on airborne targets for critical functions such as on-board alerting.
Valot, NicolasFabre, LouisPagetti, ClaireMechouche, AmmarLesage, Benjamin
This paper extends a previously developed adaptive pilot model framework for inner-loop roll-attitude tracking [1] to outer-loop position tracking tasks. Pilot Model identification is performed for two command signal types - a discrete step-like signal and a continuous Sum-of-Sines (SOS) signal - yielding distinct parameter signatures that reflect the different anticipatory and tracking demands of each signal type. An adaptive pilot model for the outer-loop position tracking task is formulated using a model-reference neural network (MRNN) architecture with a linearly parameterized neural network updated by a Lyapunov-stable adaptive law. Simulation results for both discrete and continuous tasks demonstrate that the adaptive pilot model remains stable and maintains position tracking performance under both a doubling and a halving of the nominal control sensitivity. Preliminary results are also presented for a multi-axis maritime task, extending the framework to simultaneous lateral and vertical position tracking.
Keller, AlexanderChen, ZhouzhouHorn, Joseph
This article surveys the most recent data-driven methods of lithium-ion (Li-ion) battery state of health (SOH) estimation methods and dataset resources utilized in electrified vehicles (EV) and their potential adoption for automotive battery management systems. These include regression-based models, ensemble learners, deep neural networks, and physics-informed hybrid methods. The review describes estimation methods found in articles published between 2023 and 2025, and investigates their differences in terms of estimation accuracy, data requirement, interpretability, and real-time deployment ability. The article traverses the dataset space, focusing on laboratory aging datasets, vehicle field–based datasets, telematics-derived records, and synthetic or augmented datasets, to underline that model performance in the estimation of SOH cannot be disentangled from the quality of the data, the operating coverage, and the transfer conditions. Apart from the model design, this work reviews the large-scale estimation pipeline, which involves preprocessing under sensor noise and irregular timestamps, feature extraction from incremental capacity, differential voltage, relaxation response and impedance-related indicators, and uncertainty handling for diagnostics and safety-based decision support. Practical constraints to the deployment of embedded BMS are covered. Such as ECU memory and computing limits, communication overhead, calibration effort, update approach, and functional-safety requirements. The review determines that the distance between laboratory validation and field robustness is large raising a need for more work in this area and also, that domain adaptation, federated learning, and improving benchmarking practice turn out to be promising directions for improving generalization and reproducibility. The article concludes that future advances in automotive SOH estimation will not only rely on better learning algorithms but also on improvement in the availability of realistic and field representative data, the application of robust evaluation mechanisms, and methods that are developed under real BMS constraints.
Nyachionjeka, KumbirayiBayoumi, Ehab H.E.
Accurately modeling and controlling vehicle exhaust emissions, particularly during highly transient events such as rapid acceleration, is crucial for meeting stringent environmental regulations and optimizing modern powertrain systems. While conventional data-driven modeling methods, such as Multilayer Perceptrons (MLPs) and Long Short-Term Memory (LSTM) networks, have improved upon earlier phenomenological or physics-based models, they often struggle to capture the complex nonlinear dynamics of emission formation. These monolithic architectures attempt to learn from all available data, which increases their sensitivity to dataset variability. They often require increasingly deep and complex architectures to improve performance, thereby limiting their practical utility. This paper introduces a novel approach that overcomes these limitations by modeling emission dynamics in a structured latent space. Using a rich dataset combining real-world driving data from a Portable Emission Measurement System (PEMS) with high-frequency hardware-in-the-loop test bench measurements, a Joint Embedding Predictive Architecture (JEPA) is leveraged. This framework learns to abstract away irrelevant information and encode only the key factors governing emission behavior into a compact, robust latent representation. The resulting model demonstrates superior data efficiency and predictive accuracy across diverse transient regimes, exhibiting stronger generalization than the high-performing LSTM baseline. Structured pruning and post-training quantization are applied to the JEPA framework to enhance the model’s suitability for real-world deployment. This combined strategy significantly reduces the model’s computational footprint, minimizing inference time and memory demand, with only a marginal impact on accuracy. This yields a highly accurate model well suited to on-board implementation of advanced control strategies, such as model predictive control or model-based reinforcement learning, in both conventional and hybrid electric powertrains. The results indicate a clear pathway toward more efficient and robust emission control systems for next-generation vehicles.
Sundaram, GaneshGehra, TobiasUlmen, JonasHeubaum, MirjanGörges, DanielGünthner, Michael
This paper presents a hybrid optimization framework that integrates Multi-Physics Topology Optimization (MPTO) with a Neural Network–surrogated Design of Experiments (NN-DOE) to enable lightweight structural design while satisfying crashworthiness, durability, and noise, vibration, and harshness (NVH) requirements under practical casting and packaging constraints. In the proposed MPTO formulation, crash and durability performances are incorporated through equivalent static compliance measures, while NVH performance is assessed using a frequency-domain dynamic stiffness metric, allowing consistent evaluation of trade-offs among competing design requirements. The framework is first demonstrated using a mass-produced passenger-car lower control arm (LCA) as a benchmark component. In this application, MPTO achieves weight reduction under multi-physics objectives by removing non-load-bearing material. Results show that single-discipline optimization produces unbalanced topologies, while balanced crash–durability–NVH consideration yields robust load paths. The study further demonstrates that crash and durability are dominated by static compliance–based response, whereas NVH performance is governed by frequency-dependent dynamic response over the relevant frequency range. The framework is then applied to a front engine mounting bracket of a newly developed heavy-duty truck. In this second application, a two-step strategy is employed in which MPTO first establishes the global load-carrying topology under manufacturing and packaging constraints, followed by NN-DOE–based local refinement to achieve stress attenuation at non-designable regions through global structural stiffness rebalancing, rather than direct geometric modification. Final verification confirms a steel-to-aluminum material transition achieving approximately 45% weight reduction and a substantial improvement in durability fatigue life, while maintaining required crash performance.
Kim, HyosigSenkowski, AndresGona, KiranSaroha, LalitBoraiah, Mahesh
Accurate prediction of electric vehicle charging time is critically hindered by dynamic, non-linear factors including battery aging which is indicated by the State of Health (SOH), substantial power diversion to thermal management systems in extreme temperatures, fluctuating user-defined accessory loads, and hardware limitations of the charging infrastructure. Traditional estimation methods, reliant on static models or predefined calibrations, fail to adapt to these real-world variables, leading to inaccurate predictions and user dissatisfaction. This paper presents a novel data-driven estimation framework utilizing a tailored feedforward neural network architecture specifically designed for this complex task. The model processes a sensitive set of inputs—including initial State of Charge (SOC), SOH, battery temperature, charging station power level and user-selected target SOC—to effectively capture the intricate, non-linear interdependencies governing the charging process. The network is trained offline using the Levenberg-Marquardt algorithm, which optimizes network complexity and mitigates overfitting, ensuring robust generalization without reliance on explicit electrochemical equations. A cornerstone of this invention is its continuous offline learning and update strategy; new field data from diverse charging scenarios is aggregated to periodically retrain and rigorously validate improved network parameters. These updated models are deployed seamlessly to vehicles via Flash-Over-The-Air updates, enabling the system to adapt to battery degradation and evolving usage patterns throughout the vehicle's lifespan. Validation under a wide range of conditions demonstrates a substantial increase in prediction accuracy compared to conventional model-based and calibration-based approaches. This solution, engineered for real-time deployment in vehicle control units, significantly enhances charging transparency, reliability, and overall user satisfaction by providing consistently accurate remaining charge time estimates.
Xie, ZhentaoShojaei, SinaWeslati, Feisel
This study develops a personalized driver model for expressway merging, embedding individual driving characteristics into automated longitudinal and lateral control via Long Short-Term Memory (LSTM) networks. Uniform assistance (Advanced Driver Assist System, ADAS) can feel uncomfortable when it does not match a driver’s style; we therefore target the merge maneuver—a safety-critical task requiring anticipation and timing—and test whether merging-related context improves model fidelity. Driving data were collected in a high-fidelity motion-base simulator across two merging scenarios (13 licensed drivers in total). Inputs comprised ego speed, Headway distance and relative speed to the lead vehicle, and geometric context variables (distance to the end of the acceleration lane and to the hard/soft nose); outputs were longitudinal and, in the cross-scenario study, lateral accelerations. Models were trained per driver and evaluated by root mean square error (RMSE). Including merging context reduced longitudinal error in Experiment 1 (Gotemba IC) by about 30% on average relative to models without context, while errors remained below 0.5 m/s2. In Experiment 2 (Tokyo–Nagoya Expressway vs. Tokyo Metropolitan Expressway), longitudinal and lateral errors were low across both geometries; group-mean trends favored context but were non-significant, reflecting small sample size and inter-individual variability. Questionnaire-based evaluations in the simulator showed ratings close to real driving for discomfort, merge timing, and perceived safety; similarity and willingness to use were slightly higher in the urban expressway scenario, suggesting good user acceptance in constrained conditions. These findings indicate that incorporating merging context enables personalized control that better reflects individual driving behavior, while pointing to future work on generalization across geometries, speed ranges, and richer interaction semantics.
Shen, ShuncongHirose, Toshiya
The tire model is a crucial component in the design of the K-characteristic of FSAE racing car suspensions, and directly influences the achievement of maximum cornering lateral force. Not only do the slip angle, vertical load, tire pressure, and camber angle affect the mechanical characteristics of the tire, but temperature is also an important influencing factor when FSAE vehicle tires operate at high speeds. However, the modeling process of traditional tire models based on temperature characteristics is often very complex. The FSAE tire test code (FSAE TTC) already has a large amount of official sample data, which provides a basis for data-driven neural network models. This study implemented a hybrid modeling methodology, constructing two cascaded feedforward neural networks that combine the physical interpretability of the Magic Formula tire model with the nonlinear approximation capabilities of neural networks. The first network model uses slip angle, vertical load, tire pressure, and camber angle as input features, while the second uses tire temperature, ambient temperature, and ground temperature. The first network model simulates the magic formula model of the tire, and the second fine-tunes the lateral force, aligning moment, and overturning moment based on temperature characteristics. It prevents secondary input features (such as temperature) from being completely dominated by primary input features, facilitating the explanation of the influence of the two feature groups on tire characteristics. The accuracy and robustness of the model are suitable for the engineering requirements of FSAE. During the Formula Student China competition, based on on-track measured data, the tire model was co-simulated with VI-CarRealTime to quickly calculate the tire pressure required to achieve maximum lateral force. This effectively saved practice time before the race and helped the team achieve a third-place finish.
Liu, XiyuanWang, ShenyaoLi, MingyuanHuang, Jiayu
The Formula SAE (FSAE) race track is characterized by a large number of corners, making cornering performance a key factor affecting lap time. Based on the proportional control strategy for rear-wheel steering angles, this paper proposes a steering angle optimization method using a Temporal Convolutional Network (TCN). The TCN model features a faster training speed than traditional sequential neural networks. In addition, dilated convolutions enable an exponential expansion of the receptive field without increasing computational costs, making it particularly suitable for capturing the temporal dependencies of vehicle states. By processing vehicle dynamic parameters including front-wheel steering angle, vehicle speed, yaw rate and sideslip angle, the model calculates the correction value of the rear-wheel steering angle. This correction value is then superimposed with the reference value of the rear-wheel steering angle derived from the proportional control strategy, which serves as the control value for rear-wheel steering. Rear-wheel steering can reduce the turning radius during low-speed driving and enhance the racing car’s stability during high-speed cornering. This method was validated on a typical race track via CarSim-MATLAB co-simulation, resulting in reduced lap time. To meet the real-time computing requirements of FSAE, MATLAB was used to simulate the discretization results of vehicle parameters such as vehicle speed and front-wheel steering angle, generating a Look-Up Table for rear-wheel steering angles, which provides a feasible solution for real-vehicle tests. The racing car is equipped with a manual switch for the driver to operate. The driver can manually turn the rear-wheel steering function on or off when cornering or whenever they deem it necessary.
Liu, Xiyuan
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, in particular, for the formulation of eMPC that involves NOx penalty in the cost function, online adaptation further facilitates satisfying emission constraints. In particular, both adaptive methods improve Soot limit enforcement compared to a non-adaptive controller, with the online NN providing superior performance by capturing the nonlinear dynamics of in-use changes.
Zhang, JiadiLi, XiaoKolmanovsky, IlyaTsutsumi, MunecikaNakada, Hayato
This study presents an image-derived multimodal AI framework for early-stage tire noise evaluation. The proposed model requires only multi-angle photographs captured by a standard smartphone and basic tire specifications. From these images, scaled three-dimensional (3D) meshes and fixed-view depth maps are reconstructed and combined with numerical parameters within a neural network architecture. Three input branches—a point-cloud–gradient branch, a depth-map convolutional neural network (CNN) branch, and a specification multi-layer perceptron (MLP) branch—are jointly trained using a composite loss that integrates frequency-weighted mean squared error (MSE), spectral cosine similarity, FFT-domain consistency, and A-weighted sound-level terms. A dataset of 28 tires, spanning passenger, SUV, and pickup applications for both battery electric vehicles (BEVs) and internal combustion engine (ICE) vehicles, was evaluated using leave-one-out (LOO) cross-validation. The model achieved a mean absolute error of 1.52 dB for overall A-weighted near-field noise, with a maximum error of 4.55 dB observed for a high-performance tread. Predicted 1/3-octave spectra preserved dominant peaks, valleys, and overall energy distribution, supporting engineering-level interpretation. The workflow avoids dedicated 3D scanning hardware and CAD models, enabling low-cost virtual screening prior to prototype manufacturing. In addition, the study treats near-field noise as the primary modeling target and describes a minimal, data-driven calibration pathway from near-field predictions to other exterior-noise metrics, which can be extended in future work.
Shao, GuangxinShopoff, ScottFranklin, Nicholas
Accurate torque-trace reproduction on regulatory drive cycles is central to heavy-duty diesel certification and development testing. Conventional controllers such as Proportional Integral Derivative (PID or PI) can be enhanced with gain scheduling and feedforward (FF) maps to satisfy requirements but require extensive calibration and are sensitive to nonlinearities and delay. This paper evaluates a data-driven control framework comprising a recurrent neural surrogate of engine torque (specifically an LSTM – long short-term memory) trained on engine/dynamometer data and a reinforcement learning (RL) policy trained using this surrogate (“world model”) to track requested torque while regularizing control effort. The RL policy (specifically TD3 – twin delayed deep deterministic) is benchmarked against tuned PID and PID+FF baselines on the Environmental Protection Agency’s Heavy Duty Federal Test Procedure (HD-FTP) segments using EPA regression criteria (slope, |intercept|, R2) and tracking metrics (mean absolute error - MAE, root mean square error - RMSE). TD3 reduced mean absolute error (MAE) by 58% (from 91.34 to 38.69 N·m) and root mean square error (RMSE) by 54% (from 123.5 to 56.73 N·m), improved regression to slope = .9978, |intercept| = 7.13 N·m, R2 = .9844, and cut the 95th-percentile absolute error by 60% compared to the PID+FF controller (the next best performing controller – in all categories). Results show the RL controller improves responsiveness and accuracy relative to autotuned PID+FF on the surrogate model, while reducing manual calibration effort. The approach is modular and engine-agnostic (retrain surrogate and policy) and is amenable to multi-objective extensions that incorporate emissions proxies in the reward.
Cook, JamesPuzinauskas, PauliusBittle, JoshuaHall, Spencer
To effectively improve the performance of chassis control of distributed drive intelligent electric vehicles (EVs) under difference road conditions, especially in combing road information and chassis control for improving road handling and ride comfort, is a challenging task for the distributed drive intelligent EVs. Simultaneously, inaccurate chassis control and uncertainty with system input, are always existing, e.g., varying road input or control parameters. Due to the higher fatality rate caused by variable factors, how to precisely chose and enforce the reasonable chassis control strategy of distributed drive intelligent EVs become a hot topic in both academia and industry. To issue the above mentioned, an adaptive torque vector hierarchical controller based on road level and adhesion is proposed, which optimizes the comprehensive. First, combined with the characteristic of the unbalance dynamic force caused by the air gap between the stator and the rotor of the in-wheel motor, a nonlinear vehicle model based on motor unbalanced electromagnetic force is developed. Then, using the deep neural network, an algorithm for road level and adhesion recognition based on system response data is designed. Meanwhile, an adaptive torque vector controller based on road information is designed to improve the driving safety and handling stability of chassis. Finally, the proposed algorithm is validated on the full-car test rig platform, results show that the proposed algorithm can improve chassis performance under double lane-change test. The research achievements develop a reasonable algorithm to apply to the improving road handling and ride comfort performance for distributed drive intelligent EVs.
Wang, ZhenfengZhao, GaomingZhang, ZhijieZhou, ZitaoHuang, TaishuoMa, Changye
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, internal energy, specific volume, entropy, specific heat, and mole fractions of ten primary combustion species. The model is built using TensorFlow/Keras with a three-layer feedforward architecture and optimized using the Adam optimizer. Input and output variables are normalized using standard scaling, and model performance is evaluated using MSE, MAE, and R2 metrics. Results show that the DNN achieves high accuracy across most outputs, with R2 values exceeding 0.99 for major thermodynamic variables, and provides fast inference suitable for real-time applications. This study shows that deep learning provides a viable and efficient alternative to traditional equilibrium solvers for combustion modeling, with potential applications in engine control, system-level simulations, and design optimization.
Ji, HuangchangWang, KaiGuo, ZhefengHan, YangLee, Timothy
Calibration is a major resource bottleneck and source of risk in powertrain technology development. A promising alternative to the typical design-of-experiments (DoE) approach is the use of a ‘Non-Dominated Sorting Genetic Algorithm’ (NSGA) calibration method, where an iterative process is used to directly identify the Pareto Fronts between performance metrics, for example, net mean effective pressure (NMEP) and NOx emission. The goal of the present work was to develop and demonstrate a fully ‘online’ combustion system calibration method based on an NSGA, where the algorithm operates directly on experimental data rather than empirical models as is typical in the literature. This was completed by first designing an optimal NSGA for combustion system calibration and then demonstrating its use for an experimental combustion system calibration on a single cylinder gasoline engine at one operating condition. Results from the design process here indicate that ‘online’ NSGAs have a strong potential to outperform traditional DoEs in the development of Pareto-optimal engine calibrations; however, NSGA performance is highly sensitive to the specific parameters used in the algorithm logic. The highest sensitivity was to the mutation logic within the genetic reproduction process, and second was the number of genes (calibrations) included in the overall process. Inclusion of both the Primary and Secondary non-dominated Pareto fronts in the set of Pareto-optimal calibrations was found to be critical to the success of the NSGA. When demonstrated for an experimental combustion system calibration the NSGA operated effectively and as expected, continually providing ‘upward’ pressure to generate calibrations that maximize NMEP but also ensure the breadth of the Pareto front (NOx) is scanned with high fidelity. In comparison to a traditional DoE approach, the NSGA was nearly twice as accurate in identifying calibrations along the Pareto front for the same number of total experiments. The Pareto-optimal calibrations developed by the NSGA are reasonable for these operating conditions and in excellent agreement with the literature. The present work strongly motivates and supports further development of NSGA methods for use in more complex systems and situations including for electrified and hybrid powertrains.
Mansfield, Andrew
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 scale. Additionally, we examine three strategies for modeling transient dynamics: Time-Conditional, where the temporal state is directly parameterized by time; the standard Autoregressive approach, which recursively propagates predictions through time; and a stability-enhanced Autoregressive scheme incorporating rollout-based training to improve prediction accuracy and long-term temporal consistency. The models are evaluated on a comprehensive Body-in-White (BIW) crash dataset comprising 150 detailed FE simulations using LS-DYNA. The dataset represents a structurally rich vehicle assembly with over 200 components, including 38 key components featuring variable thickness distributions to capture realistic manufacturing variability. Each model utilizes the undeformed mesh geometry and component characteristics as inputs to predict the spatiotemporal evolution of the deformed mesh during the crash sequence. Evaluation results show that the models capture the overall deformation trends with reasonable fidelity, demonstrating the feasibility of applying machine learning to structural crash dynamics. Although not yet matching full FE accuracy, the models achieve orders-of-magnitude reductions in computational cost, enabling rapid design exploration and early-stage optimization in crashworthiness evaluation.
Nabian, Mohammad AminChavare, SudeepAkhare, DeepakRanade, RishikeshCherukuri, RamTadepalli, Srinivas
The push for vehicle development through virtual prototyping and testing in motorsports highlights the critical challenge of tire model selection and calibration, especially when vehicle dynamics must be accurately captured. The calibration process for tire models such as the Pacejka Magic Formula (MF) relies on parameter identification and experimental data fitting. While optimization algorithms have been implemented to calibrate tire models, few studies explore the effects of parameter selection on overall vehicle performance, complicating prioritization for the vehicle’s modeling and simulation strategy. To bridge this gap, this paper leverages optimal control methods to quantify how the variability of MF tire model parameters propagates to the overall vehicle model and impacts lap time prediction accuracy. To achieve this, a subset of parameters critical to combined slip of the MF tire model are varied through a Design of Experiments (DOE). These variations are executed on a flat oval track to simplify the dynamics yet exhibit combined slip characteristics using a fixed vehicle configuration. The minimum lap time problem is solved using collocation methods via Dymos, an optimal control library for multidisciplinary systems. A neural network surrogate model enables an interactive profiler to visualize lap time sensitivity to tire model parameters. The primary contribution of this work is a framework that parametrically connects high-level, vehicle-wide metrics such as lap time to the calibration process and selection of tire models. The parametric and interactive nature of the framework allows high-level insights across the whole design space of tire model parameters. Insights derived from this framework provide a basis to develop a strategy for prioritizing testing and calibration efforts driven by vehicle level impacts of model parameter uncertainties.
Zarate Villazon, Angel M.Brown, IanBalchanos, MichaelMavris, Dimitri
This study presents a comparative assessment of two machine learning approaches for predicting aerodynamic drag coefficients (Cd) in automotive vehicle designs using data derived from computational fluid dynamics (CFD) simulations. The first approach employs traditional regression models trained on structured parametric data generated through controlled geometric variations, while the second approach integrates unstructured point-cloud geometry with structured metadata using a multi-modal deep learning framework. Both methods are evaluated within their respective contexts to understand their strengths, limitations and potential roles in automotive aerodynamic workflows. Rather than identifying a single best approach, the study highlights how these methods address different design needs and resource constraints, providing insights for future hybrid strategies that combine interpretability with geometric sensitivity. The work aims to establish a foundation for data-driven aerodynamic analysis in automotive design, emphasizing the role of machine learning (ML) in enhancing simulation-based engineering processes.
Kumar, GauravKhanna, Susheel
This paper explores the application of an Improved Enhanced-Boost Quasi-Z-Source Inverter in AC-connected extreme fast charging (XFC) stations for electric vehicles (EVs), aiming to reduce conversion stages and enhance system efficiency. AC-connected XFCs offer superior reliability compared to DC-connected systems due to better fault tolerance and reduced sensitivity to power fluctuations but traditionally suffer from increased complexity and reduced efficiency due to multiple conversion stages. The proposed inverter addresses this by combining DC-DC and DC-AC conversion into a single stage, simplifying the system, decreasing losses, and improving efficiency. Furthermore, this research investigates the use of Spiking Neural Networks (SNNs) for generating the precise pulse width modulation (PWM) signals required for the Quasi-Z-Source Inverter. SNNs offer potential advantages in terms of dynamic response and adaptability compared to traditional PWM techniques, allowing for optimized inverter control under varying load conditions. By operating at a higher modulation index, the inverter reduces switch stress, making it suitable for high-power applications. Analytical results and simulations demonstrate the inverter's potential, enhanced by SNN-based pulse generation, in advancing next-generation fast charging solutions by simplifying infrastructure and improving charging efficiency, positioning it as a promising technology for the future of electric vehicle charging.
Saliesh, DileepSanaboyina, PrudhviChhagar, RohnitsinghSatyanarayan, Swapna
Drivers often interact with partial automation (SAE Level 2) systems, initiating transfer of control (TOC) either by handing control over to the automation or by taking it back. Accurately predicting these interactions may inform the design of future automation systems that adapt proactively to the operating context, enhance comfort, and ultimately may improve safety. We present a context-aware framework that generates a unified driver–vehicle–environment representation by fusing data from in-cabin video of the driver and of the forward roadway with vehicle kinematics, driver glance, and hands-on-wheel behaviors. This representation was encoded in a hierarchical Graph Neural Network that classified driver-initiated TOCs to: (i) Manual-to-automation and (ii) Automation-to-manual transitions and predicted time-to-TOC. Shapley-based explainable AI was used to quantify how the importance of behavioral, contextual, and kinematic cues evolved in the seconds preceding a TOC. Analysis of a naturalistic dataset of 1,565 driver-initiated TOCs from 16 experienced drivers revealed distinct patterns. Manual-to-automation transitions were preceded by lane count increases, acceleration, and spikes in glances to the instrument-cluster. In contrast, Automation-to-manual transitions were associated with lane count reductions, higher surrounding-vehicle density, deceleration, reduction in secondary-task engagement, and higher steering wheel control. Together, these patterns highlight key cues for predicting the TOC type and time-to-TOC. Using environment-only features, the classifier achieved 78% accuracy; adding vehicle kinematics increased accuracy to 84%, and incorporating driver behavior features further improved prediction to 90%. Across prediction horizons, the Manual-to-automation TOC was consistently predicted more accurately than the automation-to-manual TOC. Shapley analyses underscore that driver behavior provided the strongest cues for predicting TOCs, highlighting the value of fusing driving context with information obtained from monitoring the driver behavior to anticipate the type of driver-automation interaction and its timing.
Zhao, ZhouqiaoGershon, Pnina
Autonomous vehicle navigation requires accurate prediction of driving path curvature to ensure smooth and safe trajectory planning. This paper presents a novel approach to curvature prediction using deep neural networks trained on GPS-derived ground truth data, rather than model predictions, providing a more accurate training signal that reflects actual vehicle motion. We develop a multi-modal neural network architecture with temporal GRU encoders that processes vision features, driver intent signals, historical curvature, and vehicle state parameters to predict curvature. A key innovation is the use of GPS-based actual curvature measurements computed from vehicle motion data (κ = ωz/v) as training supervision, enabling the model to learn from real-world driving patterns. The model is trained on 5,322 samples from real-world driving data collected on The University of Oklahoma’s Norman Campus using a Comma 3X device and a 2025 Nissan Leaf electric vehicle. Experimental results demonstrate high steering curvature prediction accuracy with a Pearson correlation coefficient of 0.805, Mean Absolute Error of 0.027654, and Root Mean Squared Error of 0.034402 on the validation set. The model achieves stable convergence within 10 epochs and maintains consistent performance across diverse driving scenarios, from straight highway segments to complex turning maneuvers. This work contributes to autonomous driving technology by demonstrating the effectiveness of GPS-supervised learning for curvature prediction, successfully deployed in OpenPilot’s production system with real-time inference at 5 Hz.
Hajnorouzali, YasamanWang, HanchenLi, TaozheBurch, CollinLee, VictoriaTan, LinArjmandzadeh, ZibaXu, Bin
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 an exceptionally low false positive rate of 0.1. Our work highlights the potential of AI and graph-based deep learning methods for advancing roundabout safety, offering a robust step toward more reliable and intelligent intersection management in the era of autonomous transportation.
Lu, DuoSatish, ManthanFarhadi, MohammadChakravarthi, BharateshYang, Yezhou
Hybrid mining trucks, as core equipment for mine transportation, face high energy consumption and significant fluctuations in power demand during cyclic operations due to prolonged exposure to demanding operating conditions characterized by heavy loads and variable working conditions. To address the issues of high energy consumption and significant fluctuations in power demand during the cyclic operation of mining trucks, this paper proposes a hybrid mining truck energy management strategy based on global SOC (State of Charge) planning and neural network optimization control. First, a powertrain model was developed for a typical operating cycle of a hybrid mining truck, and its accuracy was validated by comparing it with experimental data. Using dynamic programming algorithms to plan the SOC for single-cycle operations provides a rational reference for energy allocation across different operational phases of mining trucks during a single cycle. Next, using the powerful nonlinear mapping and self-learning capabilities of neural networks, the system learns the optimal power output sequence of the range extender during offline computation. This enables rapid decision-making and adaptive adjustment of power allocation between the range extender and battery in real-time applications, thereby effectively reducing energy consumption under cyclic operating conditions. Design a SOC tracking controller to adjust the output of the neural network optimization controller, thereby achieving tracking of the planned global SOC. Finally, validated through simulations and real vehicle tests, the proposed energy management strategy can improve fuel economy, achieving a 1.27% fuel saving compared to traditional rule-based strategies in real vehicle tests. This study provides a feasible energy management scheme for the green and efficient operation of heavy mining vehicles.
Yang, JianyuZhao, ZhiguoChen, HuiyongLi, TaoZhuang, WenyuShen, PeihongTang, Peng
Fused filament fabrication (FFF) has gained popularity in recent years because it can produce prototypes and functional components with complex geometry. Because of inherent process variability, the components often exhibit defects such as warping, layer delamination, voids, and poor surface finish, as well as issues related to variable material strength and anisotropy. In-situ monitoring (ISM) of the FFF process is a promising technique to predict part performance, which in turn can support accept or reject decisions for printed parts. This paper proposes a framework for incorporating ISM-generated information, with a particular focus on infrared (IR) image analysis for this purpose. IR camera images, in conjunction with numerical features such as infill pattern and extruder nozzle temperature, serve as an input to a multimodal deep learning (MDL) model that predicts the mechanical performance of printed parts. In the framework, convolutional neural nets process image inputs, while a fully connected neural network extracts patterns from numerical process parameters. Furthermore, the proposed approach incorporates an ablation study and Cohort Shapley analysis to identify the most informative monitoring modalities and process parameters. This fusion of modalities enables more accurate and robust prediction of mechanical response than a single-source model. We demonstrate the framework on FFF-printed beams subjected to torque and three point bending tests, and discuss opportunities for future work in vehicle manufacturing and expeditionary sustainment.
Mollan, CalahanKulkarni, SaurabhMalik, Ali AhmadPatterson, Albert E.Pandey, Vijitashwa
Items per page:
1 – 50 of 1486