⚡ electrical engineering

Automotive Engineering-Centric Agentic AI Workflow Framework

This paper introduces Agentic Engineering Intelligence (AEI), a framework that models automotive engineering workflows as constrained, history-aware sequential decision processes where AI agents support engineer-supervised interventions across toolchains, demonstrated through diverse use cases like suspension design and MBSE to advance process-level industrial intelligence.

Original authors: Tong Duy Son, Zhihao Liu, Piero Brigida, Yerlan Akhmetov, Gurudevan Devarajan, Kai Liu, Ajinkya Bhave

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tong Duy Son, Zhihao Liu, Piero Brigida, Yerlan Akhmetov, Gurudevan Devarajan, Kai Liu, Ajinkya Bhave

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build the perfect car. You have a team of brilliant engineers, but they are stuck in a frustrating cycle: they design a part, run a computer simulation, the simulation fails, they stare at error logs for hours, guess what went wrong, tweak the design, and repeat. Often, they make the same mistakes they made on a project five years ago because that knowledge was lost in a pile of old PDFs and forgotten meeting notes.

This paper proposes a new way to work called Agentic Engineering Intelligence (AEI). Think of AEI not as a robot that replaces engineers, but as a super-smart, hyper-organized co-pilot that helps engineers navigate the entire journey of building a car, not just the individual steps.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Island" Approach

Currently, engineering tools work like islands. You have a tool for designing the shape, a different tool for testing the engine, and another for checking safety. When something goes wrong, the tools don't talk to each other. The engineer has to be the bridge, manually connecting the dots, remembering past failures, and guessing the next step. It's like trying to solve a puzzle where the pieces are in different rooms, and you have to run back and forth to find them.

2. The Solution: The "Conductor" (The Agent)

AEI introduces an AI Agent that acts like a conductor of an orchestra.

The Orchestra: The various engineering tools (simulations, CAD software, databases).
The Music: The goal of building a safe, fast, efficient car.
The Conductor (AI): It doesn't play the instruments itself. Instead, it listens to the whole room, remembers the sheet music from previous concerts, and tells the engineers exactly which instrument to tune next to fix a bad note.

3. How the "Conductor" Learns (The Memory)

One of the biggest problems in engineering is that knowledge disappears. If an engineer figures out why a suspension part failed in 2020, that lesson often vanishes by 2024.

AEI fixes this by building a Massive, Smart Library:

The Old Way: You search for "suspension failure" and get a list of text documents.
The AEI Way: The AI reads the text, looks at the charts, listens to the audio recordings of engineers explaining why they made certain choices, and connects the dots. It knows that a specific squiggly line on a graph means "overheating" only when combined with a specific engine setting mentioned in a meeting recording. It turns messy, scattered data into a clear, searchable story.

4. The "Energy" Meter (Making Decisions)

When the AI sees a problem (like a car part that doesn't fit), it doesn't just guess. It uses a concept called "Workflow Energy."

Imagine you are hiking up a mountain (the goal is the peak).

Performance Gap: How far are you from the top?
Constraint Violation: Are you hitting a cliff or a swamp?
Cost: How much energy does it take to climb this path?

The AI calculates the "energy" of every possible next step. It asks: "If we change this bolt, does it get us closer to the top without hitting a cliff, and is it worth the effort?" It ranks the options and says, "Try this one first," or "This path is a dead end, let's try a different route."

5. Real-World Examples

The paper shows how this works in five different scenarios:

Suspension Design: Instead of the engineer manually tweaking 50 different numbers until the car handles well, the AI looks at past failures, says, "Hey, we tried this before and it broke because of the tire angle. Let's try adjusting the wheel angle instead," and suggests the fix.
AI Tuning: Training an AI to drive a car is like tuning a radio with 100 knobs. The AEI agent remembers which knob combinations worked last time and guides the engineer to the sweet spot faster.
Aerodynamics (Wind): Testing how wind flows over a car usually takes hours of supercomputer time. The AI uses a "crystal ball" (a fast, simplified model) to test 100 ideas in seconds, picks the best 3, and then tells the supercomputer to do the detailed, expensive work only on those winners.
MBSE (The Big Picture): If a requirement changes (e.g., "The battery must stay cooler"), the AI instantly traces how that change ripples through the whole car design, warning the engineer, "If you cool the battery, the cabin might get too hot. Here are three ways to fix that trade-off."

6. The Human Role: The Captain

The most important part of this paper is that the AI does not take the wheel. It is a "Human-in-the-Loop" system.

The AI is the Navigator: It maps the terrain, points out the potholes, and suggests the best route.
The Engineer is the Captain: They make the final decision. They look at the AI's suggestion and say, "Good idea, let's do it," or "No, that won't work for safety reasons, try something else."

Summary

This paper is about moving from doing tasks to managing the process. It's about giving engineers a "second brain" that remembers everything, connects the dots between different tools, and helps them make better decisions faster. It turns engineering from a game of "guess and check" into a guided, intelligent journey toward the perfect car.

1. Problem Statement

Current engineering workflows in the automotive sector (e.g., suspension design, control tuning, aerodynamic optimization, and Model-Based Systems Engineering) are inherently iterative, multi-objective, and constraint-driven. However, existing AI and optimization methods suffer from three critical limitations:

Task Isolation: AI treats engineering activities as isolated tasks (e.g., parameter optimization or prediction) rather than parts of a broader, sequential decision-making process.
Lack of History Awareness: Engineering workflows rely heavily on expert intuition because historical data (simulation logs, failure cases, design iterations) is rarely used systematically. This leads to repeated trial-and-error cycles and the loss of institutional knowledge.
Rigid Pipelines: Conventional pipelines fail to adapt when optimization stalls. They lack mechanisms to suggest alternative interventions (e.g., revising constraints, reformulating objectives) or to diagnose root causes of infeasibility automatically.

The core problem is the need to shift from artifact-level optimization (optimizing a single design) to process-level intelligence (optimizing the engineering workflow itself).

2. Methodology: Agentic Engineering Intelligence (AEI)

The authors propose Agentic Engineering Intelligence (AEI), a framework that models engineering workflows as constrained, history-aware sequential decision processes. The framework operates in two coupled phases:

A. Theoretical Formulation

The workflow is mathematically defined as $W = (S, O, A, Z, C, T, U)$ , where:

$S$ : Latent workflow states (artifacts, requirements, logs).
$O$ : Observations (tool outputs, reports, retrieved context).
$A$ : Candidate interventions (actions).
$Z$ : Retrieved workflow memory.
$C$ : Feasibility and specification constraints.
$T$ : Transition dynamics.
$U$ : Utility model.

Control-Theoretic Interpretation:

Reference Signal: Engineering objectives.
Controller: The AI Agent.
Plant: The engineering toolchain.
Feedback: Toolchain outputs used to select interventions.

Workflow-Energy Heuristic:
To guide decision-making in partially observed environments, AEI introduces a "workflow-energy" metric ( $E_t$ ) to rank candidate actions:
$E_t(a) = \alpha \hat{G}_t(a) + \beta \hat{V}_t(a) + \gamma \hat{C}_t(a)$
Where:

$\hat{G}_t$ : Estimated remaining performance gap.
$\hat{V}_t$ : Estimated residual constraint violation.
$\hat{C}_t$ : Estimated workflow cost.
The agent selects the action $a^*$ that minimizes this energy, serving as a stability-inspired guidance signal rather than a formal proof of stability.

B. Technical Architecture

Offline Phase (Memory Construction):
- Transforms heterogeneous engineering data (CAD, simulation logs, reports, slide decks) into a structured Engineering Memory ( $D$ ).
- Multimodal Processing Pipeline: Uses a three-path approach to handle complex engineering data:
  - Text/Table Extraction: For semantic and keyword search.
  - Context-Aware Figure Interpretation: Uses Vision Language Models (VLMs) with surrounding text and narration to interpret engineering figures (e.g., contour plots) physically, not just visually.
  - Speech-to-Text: Aligns oral explanations from presentations with specific slides to capture tacit knowledge (e.g., "why a mesh density was chosen").
- Hybrid Retrieval: Combines vector databases (semantic similarity) with keyword-indexed databases (exact parameter matches) to support both conceptual and precise technical queries.
Online Phase (Closed-Loop Planning & Control):
- The agent estimates the current workflow state, retrieves relevant history ( $Z$ ), reasons over constraints, and ranks interventions.
- Human-in-the-Loop: The system is recommendation-centered. It proposes ranked actions (e.g., relax a constraint, restart optimization, escalate to human), but the engineer retains final decision authority.

3. Key Contributions

Framework Proposal: Introduces AEI as an industrial vision framework that recasts engineering workflows as sequential decision processes, shifting focus from optimizing artifacts to optimizing the process.
Compact Formulation: Provides a unified mathematical abstraction for workflow state, memory retrieval, constraints, and utility, applicable across diverse engineering domains.
Control-Theoretic Lens: Interprets agentic AI as a controller for engineering toolchains and introduces the Workflow-Energy Heuristic for ranking interventions based on performance gaps, constraint violations, and cost.
Multimodal Knowledge Reuse: Develops a pipeline to convert unstructured engineering artifacts (including oral rationale and figures) into searchable, context-rich memory.
Empirical Validation: Demonstrates the framework across five distinct automotive use cases.

4. Results and Use Cases

The paper validates the AEI framework through five representative automotive scenarios:

Suspension Design (Target Cascading):
- Challenge: Translating Kinematics & Compliance (K&C) targets into 3D hardpoint configurations often leads to infeasible designs due to conflicting packaging constraints.
- AEI Solution: The agent treats failed optimization runs as informative states. It analyzes violation patterns, retrieves similar past cases, and recommends specific interventions (e.g., relaxing bounds, adjusting initialization) rather than simply stopping.
Reinforcement Learning (RL) Control Tuning:
- Challenge: Tuning hyperparameters for RL controllers is non-linear and relies on tacit expert knowledge.
- AEI Solution: The agent combines a human-authored knowledge base with an agent-built memory of training runs. It interprets training dynamics to recommend targeted hyperparameter adjustments, creating a traceable log of tuning decisions.
Multimodal Engineering Knowledge Reuse:
- Challenge: Critical design rationale is often buried in figures or oral presentations, making it inaccessible to standard text-based search.
- AEI Solution: The multimodal pipeline successfully extracts and links visual, textual, and spoken data, allowing the agent to retrieve "why" a decision was made, not just "what" the result was.
Surrogate-Assisted Aerodynamic Design:
- Challenge: High-fidelity CFD simulations are too slow for broad design exploration.
- AEI Solution: The agent uses AI surrogates for rapid screening but employs workflow-level reasoning to decide when to escalate to high-fidelity CFD. It balances prediction speed with the need for verification.
MBSE Integration (Simcenter):
- Challenge: Managing requirement conflicts across subsystems (e.g., thermal management vs. cabin cooling).
- AEI Solution: The agent performs Sense-Reason-Act cycles: sensing requirement changes via traceability links, reasoning about cross-domain trade-offs using historical data, and proposing specific workflow interventions (e.g., retuning control logic) to the engineer.

5. Significance

Process-Level Intelligence: The paper fundamentally shifts the paradigm of Engineering AI from "solving a specific problem" to "managing the engineering process," enabling adaptive, history-aware decision support.
Bridging Tacit and Explicit Knowledge: By capturing oral rationale and figure context, AEI addresses the critical gap where institutional knowledge is lost when teams change or projects end.
Practical Roadmap: Unlike theoretical autonomous systems, AEI is designed as a recommendation-centered tool that augments human engineers, ensuring safety and traceability in industrial settings.
Scalability: The framework offers a common abstraction that can be applied across diverse domains (mechanical, control, systems engineering), suggesting a path toward unified industrial AI ecosystems.

The paper concludes that while AEI does not claim full autonomy, it provides a robust, practical roadmap for integrating agentic AI into industrial engineering workflows, with future work focusing on empirical validation and tighter integration with traceable toolchains.