[Paper] Data-Driven Methods and AI in Engineering Design: A Systematic Literature Review Focusing on Challenges and Opportunities
Published: (November 25, 2025 at 06:16 AM EST)
4 min read
Source: arXiv
Source: arXiv - 2511.20730v1
Overview
A new systematic literature review maps how data‑driven methods (DDMs) and AI are being used across the engineering design lifecycle. By analysing 114 peer‑reviewed studies from the past decade, the authors expose where machine‑learning (ML) techniques are thriving, where they’re still rare, and what hurdles engineers face when trying to embed AI into real‑world product development.
Key Contributions
- Comprehensive mapping of DDMs (ML, statistical, deep learning, surrogate models) onto the four stages of the V‑model: system design, implementation, integration, and validation.
- Quantitative trends showing ML and classic statistical tools dominate today, while deep learning usage is accelerating.
- Identification of gaps: validation stage receives far fewer AI‑driven contributions, and cross‑stage traceability remains weak.
- Challenge taxonomy covering interpretability, data quality, model transferability, and real‑world validation.
- Road‑map for future work, calling for interpretable hybrid models and tighter alignment between computer‑science algorithms and engineering design tasks.
Methodology
The authors followed the PRISMA systematic review protocol:
- Scope definition – Adopted a simplified V‑model (design → implementation → integration → validation) as the reference framework.
- Database search – Queried Scopus, Web of Science, and IEEE Xplore for papers published between 2014‑2024 using keywords around “data‑driven”, “AI”, and “engineering design”.
- Screening – From an initial 1,689 records, duplicates and out‑of‑scope papers were removed, leaving 114 studies for full‑text analysis.
- Classification – Each study was coded for:
- Type of DDM (e.g., supervised learning, clustering, deep learning, surrogate modeling)
- Lifecycle stage where it was applied
- Reported challenges and validation approach
- Synthesis – Aggregated counts, trend lines, and thematic analysis produced the final insights.
The process is deliberately transparent so that other researchers can replicate or extend the review.
Results & Findings
| Lifecycle Stage | Dominant DDMs | Emerging Techniques | Notable Gaps |
|---|---|---|---|
| System Design | Supervised learning (regression, classification), clustering, surrogate models | Deep learning (e.g., generative design) – still <10% | Limited use of reinforcement learning for concept exploration |
| System Implementation | Regression, statistical DOE, surrogate modeling | DL for component‑level prediction | Few studies address real‑time model updates |
| System Integration | Multi‑objective optimization, clustering, surrogate models | DL for system‑level performance prediction | Sparse coverage of cross‑disciplinary data fusion |
| Validation | Mostly statistical validation (cross‑validation, error metrics) | Almost no DL‑based validation | Lack of field‑testing or hardware‑in‑the‑loop experiments |
- Trend: Deep learning citations grew from <2 % in 2014 to >15 % in 2023, indicating rising confidence but still early‑stage adoption.
- Challenges:
- Interpretability: Engineers struggle to trust black‑box models for safety‑critical decisions.
- Traceability: Linking a model’s output back to design requirements across stages is cumbersome.
- Real‑world validation: Most papers stop at simulated validation; few deploy models on physical prototypes.
Practical Implications
- For Product Development Teams – Off‑the‑shelf ML tools (regression, clustering) are mature enough for early‑stage design and integration tasks. Teams can start small, using these methods to accelerate trade‑off studies without heavy infrastructure.
- Tool Vendors – There’s a market opportunity for platforms that embed interpretability (e.g., SHAP, LIME) and version‑controlled model provenance directly into PLM/ALM systems, closing the traceability gap.
- AI Engineers – The upward trend in deep learning suggests a need to develop domain‑specific architectures (e.g., graph neural nets for CAD geometry) that can be safely transferred to later stages.
- Quality & Safety Assurance – The scarcity of validation‑stage AI calls for new testing frameworks that combine simulation, digital twins, and hardware‑in‑the‑loop experiments, enabling regulators and certifiers to assess AI‑augmented designs.
- Education & Training – Curriculum designers should emphasize hybrid modeling (combining physics‑based and data‑driven components) to produce engineers who can both build and critically evaluate AI models.
Limitations & Future Work
- Scope restriction – The review only considered papers indexed in three major databases; relevant industry white‑papers or conference demos may be missing.
- V‑model simplification – Real‑world development often follows iterative or agile cycles, which the four‑stage mapping may not fully capture.
- Depth of analysis – While the study quantifies method prevalence, it does not evaluate the comparative performance of different DDMs on identical design problems.
Future research should (1) create a taxonomy that directly links specific AI algorithms to concrete engineering design tasks, (2) develop guidelines for building interpretable hybrid models, and (3) design robust validation pipelines that bring AI from simulation into physical prototypes.
Authors
- Nehal Afifi
- Christoph Wittig
- Lukas Paehler
- Andreas Lindenmann
- Kai Wolter
- Felix Leitenberger
- Melih Dogru
- Patric Grauberger
- Tobias Düser
- Albert Albers
- Sven Matthiesen
Paper Information
- arXiv ID: 2511.20730v1
- Categories: cs.SE, cs.AI, cs.LG
- Published: November 25, 2025