On This Page
- The Real Story Isn't Only Self-Driving Cars
- Which AI Path Should You Choose?
- Design Stage: Generative Design
- Manufacturing Stage: Digital Twins
- In-Service Stage: Predictive Maintenance
- A Fourth Layer: ADAS Perception
- The Vehicle Lifecycle, Mapped
- Tool Matrix & Projectium Verdict
- What AI Cannot Prove
- FAQ
The Real Story Isn't Only Self-Driving Cars
Most students, when questioned about the role of AI in automobile engineering, will almost always answer you with autonomous driving. That's not bad, but it's not a complete solution and can result in a common error: selecting an AI tool based on popularity rather than suitability for a specific problem. As the name suggests, AI should be used where it is needed in the vehicle lifecycle, not merely because it's a popular AI tool. A technique for a component in the design phase, a subsystem that is awaiting validation and a vehicle on the road will all need a different technique and will have a different level of student feasibility. It's presented around that life cycle, and it's truthful about what each stage entails.
Which AI Path Should You Choose?
Match The Lifecycle Stage To Your Project, Not The Other Way Around
Design Stage: Generative Design
Generative design tools produce several manufacturable candidate designs based on prescribed loads, constraints, and manufacturing processes, instead of the engineer having to make trial-and-error design choices.
In 2025, a sports vehicle steering wheel was used as a case study to apply generative design in the Autodesk Fusion 360 software, setting load cases and manufacturing constraints for additive manufacturing. The final design was optimised to achieve a 96% mass reduction from the reference aluminium wheel, and an eight-fold safety margin was demonstrated using physical tests[1].
Create a realistic load case for a real automotive component in Fusion 360's generative design module (free education license), and generate weight-optimised geometry based on that load case. There is no requirement for advanced optimisation theory knowledge to obtain a useful result, and most other mechanical engineering project ideas involving structural parts follow the same workflow.
Optimal structures in simulation are not necessarily manufacturable. Even with support removal, tolerances, and the type of additive or subtractive process, a design-for-manufacture review is still required before it becomes anything more than software.
Manufacturing Stage: Digital Twins
A digital twin for a tracked vehicle was documented in a technical paper in 2024 SAE and compared the on-board sensor and video data from real test scenarios with the digital twin's predictions, which proved it to be a true predictive tool for maintenance planning and design studies and not just a visualisation exercise[2].
Attach one subsystem (e.g., a suspension model or a small motor test rig) with sensors, and develop a simulation model that captures the real behaviour of the subsystem. The same process of simulation and validation that is discussed in our guide to the use of AI tools for industrial engineering guide is discussed here. Allow budgeted time to run the rig under conditions that were not included in the model construction, and compare predicted and actual behaviours when working with the unseen data.
A model that has only been trained with the same conditions that it was tested on has failed to show predictive ability. Knowing the accuracy of the simulation in known scenarios does not constitute physical validation, and this is a question that examiners who deal with simulation tend to ask directly.
In-Service Stage: Predictive Maintenance
The study used real EV sensor data to evaluate the performance of the Random Forest, Decision Tree, Logistic Regression, and Neural Network models, concluding that both Random Forest and Neural Networks achieved an accuracy of 96.6 7% in predicting failure mode[3].
Apply public OBD-II or EV sensor datasets to classify using scikit-learn classification models. The most approachable of the four areas: standard tabular classification, no specialised computer vision or generative design software needed, and it fits well with the EV-related topics in our EEE final year project ideas list.
The published benchmark accuracy of 96.67 % is only for this specific dataset; it does not necessarily mean the same accuracy will be obtained in the real world in cases of vehicle age, driving conditions, or sensor drift that were not covered by this dataset. There is a difference between measuring model accuracy and real-world reliability.
A Fourth Layer: ADAS Perception
A detailed 2025 survey of 3D object detection techniques for autonomous driving showed the evolution of the field from single-modal deep feature extraction to multi-sensor fusion, where maintaining the accuracy of object detection while meeting real-time processing constraints continues to be an active and unsolved challenge[4].
Tune an off-the-shelf detection model (YOLO) to a single, narrow hazard-detection task on KITTI or nuScenes instead of a complete perception stack. Combining camera data with one additional sensor, radar or LiDAR, gives a more realistic idea of production systems than camera-only detection, and the model-tuning process is very similar to the AI tools for computer science students.
The benchmark of object detection accuracy cannot be used as an indicator of the circumstances around the object. Object detection is not synonymous with autonomous-driving safety, which requires much more than a model that detects objects correctly in test conditions.
The Vehicle Lifecycle, Mapped
Fig. 1: Where Each AI Application Sits Across The Vehicle Lifecycle
Unlike the other three, ADAS perception is always with us, influencing design decisions, confirmed in the digital twin before entering the field, and monitored in service. That's one of the reasons it's the most complicated of the four to work on by itself.
Tool Matrix
| Sr. No. | Lifecycle Stage | Recommended Tool | Realistic Student Scope |
|---|---|---|---|
| 1 | Design | Fusion 360 Generative Design | Redesign one existing component under defined load cases |
| 2 | Manufacturing | Simio, or a custom Python simulation | Digital twin of a single subsystem, not a full vehicle |
| 3 | In-Service | scikit-learn (Random Forest, NN) | Failure prediction on public or OBD-II sensor data |
| 4 | Perception | YOLO fine-tuning on KITTI/nuScenes | One narrow detection task, not a full perception stack |
Predictive maintenance is the most feasible starting point for a typical 4th-year automotive engineering student, not because the student wants to, but because it is feasible. It does not require any lab equipment, uses public OBD-II or EV sensor datasets, relies on standard scikit-learn classifiers, and has a published and replicable benchmark for comparison[3].
The physical rig and disciplined usage of unseen data are required for digital twin projects; CAD-FEA fluency and access to licensed software are needed for generative design; and GPU compute and large annotated data are required for ADAS perception. They're all better projects in theory, but they're all more likely to fail given a one-semester timeline. The honest bargain: predictive maintenance is a beaten path, so a project here is a winner if it is executed and if the amount of analysis done is thorough.
What AI Cannot Prove
- Simulation result ≠ physical validation. A digital twin that comes with its own training conditions has not yet demonstrated anything about the actual behaviour of the system under conditions not demonstrated.
- Model accuracy ≠ real-world reliability. A published accuracy figure does not necessarily reflect the performance in all driving conditions, under all vehicle ages, under all sensor states, but only in one set of conditions.
- Generated geometry ≠ automatically manufacturable design. Design for manufacture review is still required for structural optimality to become a part of the software.
- Object detection ≠ autonomous-driving safety. Correct detection is a component of the safety case; it is not the safety case itself.
Frequently Asked Questions
Choose the lifecycle stage in which the equipment and data you are interested in are actually available this semester, rather than the most advanced lifecycle stage. A good use of predictive maintenance from a public data source will present themselves well in a viva voce examination than an aspiring perception project that is never reliable on real hardware.
References
- Employing generative design methods to develop an optimized automotive part (2025). ScienceDirect. Supports: Design Stage section. View source · Back to text
- Daly, N., Manvi, P., Chhatbar, T., Schmid, M., et al. (2024). Modeling & Validation of a Digital Twin Tracked Vehicle. SAE WCX World Congress Experience. DOI: 10.4271/2024-01-2323. Supports: Manufacturing Stage section. View source · Back to text
- Improving Electric Vehicle Maintenance by Advanced Prediction of Failure Modes Using Machine Learning Classifications (2025). Eksploatacja i Niezawodność (Maintenance and Reliability). Supports: In-Service Stage section and Projectium Verdict. View source · Back to text
- Three-dimensional object detection for autonomous driving via deep learning: A review (2025). ScienceDirect. Supports: ADAS Perception section. View source · Back to text