Engineering AI: Automating Complex Automotive Workflows

Author: Denis Avetisyan

A new framework models engineering processes as sequential decisions, enabling AI agents to tackle intricate challenges in automotive development.

Agentic AI integrates with engineering toolchains to perceive workflow status from existing models and analyses, deliberate upon potential actions, and facilitate workflow execution under engineer oversight-a system designed not by construction, but by anticipating and accommodating inevitable failure states.

This paper introduces Agentic Engineering Intelligence (AEI), a system leveraging memory, constraint handling, and human oversight to optimize engineering workflows modeled as sequential decision processes.

Iterative engineering workflows-spanning design optimization, simulation, and model-based systems engineering-are often treated as isolated tasks despite being fundamentally shaped by historical decisions and constraints. This paper introduces the ‘Automotive Engineering-Centric Agentic AI Workflow Framework’ and proposes Agentic Engineering Intelligence (AEI), which models these workflows as constrained, history-aware sequential decision processes supported by AI agents acting under engineer supervision. By linking offline data processing with online workflow-state estimation and decision support, AEI offers a unified approach to leverage memory and enhance intervention selection within complex toolchains. Could this framework represent a shift toward process-level intelligence, enabling more efficient and robust engineering AI systems in automotive and beyond?

The Inevitable Complexity: Why Engineering Demands Intelligent Systems

Conventional engineering design processes frequently rely on repeated cycles of prototyping, testing, and refinement, a methodology inherently burdened by substantial manual intervention. This iterative approach, while often necessary, introduces considerable delays as engineers meticulously analyze results and implement changes by hand. The accumulation of these manual tasks not only extends project timelines but also significantly elevates associated costs, impacting resource allocation and overall project budgets. Consequently, even seemingly minor design adjustments can necessitate extensive rework, hindering innovation and potentially compromising the competitiveness of engineered products. The inherent inefficiencies of these traditional workflows underscore the pressing need for more streamlined, automated solutions capable of accelerating the design process and reducing financial burdens.

Contemporary engineering challenges increasingly involve systems of such intricate interdependence that traditional design and optimization methods struggle to deliver efficient solutions. The sheer number of variables and potential interactions within modern infrastructure, from aerospace components to microchip fabrication, overwhelms manual processes and necessitates a move towards automated intelligence. This demand isn’t merely about speed; it’s about exploring a design space too vast for human comprehension, identifying subtle performance bottlenecks, and proactively mitigating potential failures. Consequently, researchers are focusing on developing algorithms and computational frameworks capable of autonomously navigating complex design landscapes, learning from simulations and real-world data, and ultimately, creating more robust, efficient, and innovative engineered systems.

The escalating intricacy of modern engineering systems necessitates a paradigm shift from conventional, iterative design processes to those modeled as sequential decision-making procedures. This work addresses this need through the introduction of Agentic Engineering Intelligence, a framework which conceptualizes engineering workflows not as static sequences, but as a series of interconnected choices made by autonomous ‘agents’. By representing each design task or optimization step as a decision point, the framework allows for the application of artificial intelligence techniques – specifically, reinforcement learning – to navigate the design space efficiently. This agent-based approach facilitates automated exploration of potential solutions, adaptive refinement of designs based on performance feedback, and ultimately, a significant reduction in both development time and associated costs, promising a future where complex engineering challenges are tackled with intelligent, adaptive workflows.

The Agentic Engineering Intelligence (AEI) framework integrates offline data processing and memory construction with online closed-loop planning and control to enable agents to reason over engineering workflows, leveraging prior experience and constraints across diverse use cases like suspension design, aerodynamic optimization, and model-based systems engineering.

The Ghost in the Machine: Formalizing Engineering as Intelligent Action

Agentic Engineering Intelligence (AEI) formalizes engineering workflows as sequential decision processes subject to inherent constraints – such as physical laws, design requirements, and budgetary limitations. This approach represents a workflow as a series of discrete actions an agent can take to move closer to a defined goal state. Crucially, AEI incorporates a historical awareness, allowing the system to learn from previous actions and their outcomes. By framing engineering problems in this manner, AEI enables the development of automated problem-solving capabilities, where an agent can autonomously navigate the design space, evaluate potential solutions, and iteratively refine its approach based on feedback and accumulated experience. This contrasts with traditional methods by explicitly modeling the process of engineering, rather than solely focusing on the final design output.

Offline Memory Construction within the Agentic Engineering Intelligence framework involves the systematic recording and indexing of past problem-solving attempts, including both successful and unsuccessful strategies, along with associated contextual data such as design parameters and environmental conditions. This historical data is then utilized to build a searchable knowledge base that allows the agent to rapidly assess the potential efficacy of new solutions by identifying analogous situations encountered previously. By referencing this memory, the agent can prioritize promising approaches, avoid known failure modes, and accelerate the optimization process, effectively reducing redundant exploration and improving overall problem-solving efficiency.

Agentic Engineering Intelligence builds upon established Model-Based Systems Engineering (MBSE) methodologies by incorporating agentic decision-making and historical experience. While MBSE provides a structured approach to defining and analyzing system requirements and designs, this framework enhances it with automated problem-solving capabilities and adaptability. Specifically, the introduced framework moves beyond static model analysis by enabling dynamic solution exploration and refinement based on past performance, resulting in a more robust response to complex and evolving engineering challenges. This work demonstrates the framework’s capabilities and establishes its potential to improve efficiency and effectiveness in various engineering applications.

This agentic suspension design workflow incorporates human supervision by monitoring optimization results, geometric constraints, and past designs to recommend targeted actions during hardpoint design.

Closing the Loop: The Illusion of Control

Online Closed-Loop Planning operates by continuously assessing the present condition of a workflow, leveraging data from current state estimation. This system doesn’t rely on pre-programmed sequences; instead, it dynamically selects interventions based on a combination of real-time workflow status, previously successful strategies stored in retrieved memory, and predefined operational constraints. The process involves evaluating potential actions and choosing the optimal one to achieve desired outcomes, adapting to changing conditions as they occur and enabling continuous optimization without manual rescheduling or intervention. This contrasts with open-loop systems which lack the feedback mechanism for real-time adjustment.

Workflow State Estimation relies on the continuous collection and analysis of data points representing the engineering process, including task completion, resource allocation, and identified anomalies. This data is then processed using techniques like Bayesian networks and Kalman filtering to construct a probabilistic model of the current workflow condition. The resulting state estimation provides a quantified assessment of progress, potential bottlenecks, and deviations from the planned trajectory, which serves as a critical input for subsequent decision-making processes within the closed-loop planning system. Accurate state estimation is essential for effective intervention selection and optimization, as it allows the system to respond dynamically to changing conditions and proactively address potential issues.

Intervention Selection utilizes the Workflow-Energy Heuristic, a methodology designed to optimize actions within an engineering workflow based on both effectiveness and resource consumption. This heuristic assesses potential interventions by quantifying their anticipated impact on workflow progression – maximizing performance metrics – while simultaneously calculating associated costs, including computational expense, time delays, and resource allocation. The selection process prioritizes interventions exhibiting the highest ratio of projected performance gain to total cost, thereby ensuring efficient resource utilization and minimizing overall expenditure. This approach allows for dynamic adaptation to changing workflow conditions and constraints, enabling real-time optimization of engineering processes.

This workflow leverages rapid surrogate predictions to guide geometry optimization, selectively employing high-fidelity CFD analysis only when necessary to balance computational cost and design performance.

The Echo of Simulation: Validation and the Promise of Predictability

Amesim and Simcenter are utilized as core components in the validation and refinement of the decision-making framework through comprehensive modeling and simulation. These platforms allow for the creation of virtual prototypes and the execution of numerous simulations under varied operating conditions, enabling the assessment of framework performance without physical testing. Specifically, Amesim specializes in the modeling of multi-domain physical systems, while Simcenter offers advanced capabilities in structural, thermal, and fluid dynamics analysis. By comparing simulation results against expected outcomes or experimental data, engineers can identify and correct deficiencies in the framework’s algorithms and parameters, ultimately increasing confidence in its reliability and accuracy. This iterative process of simulation, analysis, and refinement is critical for ensuring the framework consistently delivers optimal decisions across a range of applications.

Surrogate models, leveraging the PhysicsAI engine, address computational bottlenecks inherent in complex system design optimization. Traditional physics-based simulations, while accurate, can be excessively time-consuming when evaluating numerous design iterations. PhysicsAI creates data-driven, reduced-order models – the surrogate models – trained on a limited set of high-fidelity simulation results. These surrogates then rapidly predict system behavior for new design inputs, reducing evaluation times from hours or days to seconds or minutes. This acceleration is critical for iterative optimization loops, enabling engineers to explore a wider design space and identify optimal solutions more efficiently. The accuracy of the surrogate model is dependent on the quality and quantity of the training data, and PhysicsAI incorporates techniques to quantify and minimize prediction uncertainty.

Teamcenter serves as a product lifecycle management (PLM) system centralizing all engineering data and artifacts, including CAD models, requirements documents, simulation results, and testing data. This centralized repository enables controlled access and versioning, preventing data loss and ensuring all team members utilize current, validated information. Specifically, Teamcenter supports workflow and change management, facilitating collaborative design reviews and approval processes. The system’s capabilities extend to managing bill of materials (BOMs) and configuration data, critical for complex system development. By providing a single source of truth for engineering information, Teamcenter minimizes errors, reduces time-to-market, and streamlines the overall development lifecycle.

An engineering knowledge processing pipeline integrates data from diverse sources-slide decks, reports, and recordings-by leveraging a vision-language model with contextual text and narration to create physically grounded descriptions of figures, and then indexes this information in a hybrid semantic/keyword store for use by an agent during decision-making.

The Inevitable Future: A Paradigm Shift in Engineering Practice

Agentic Engineering Intelligence presents a transformative approach to automotive design, particularly in the areas of suspension and aerodynamics. This innovative framework moves beyond traditional optimization techniques by enabling artificial intelligence to autonomously formulate, execute, and refine engineering solutions. Studies indicate that this capability drastically reduces development cycles, as the system independently explores design spaces and identifies high-performing configurations that might elude human engineers. The resulting vehicles demonstrate marked improvements in handling, stability, and fuel efficiency, achieved through designs tailored to specific performance criteria. By automating complex tasks and fostering a continuous cycle of learning and improvement, this technology promises to redefine the future of automotive engineering and unlock new levels of vehicle performance.

Beyond automotive engineering, the agentic framework exhibits remarkable potential across a spectrum of complex design challenges. Its core principles – leveraging historical data, intelligent planning, and iterative refinement – are readily transferable to fields like aerospace, biomedical device development, and even architectural design. This adaptability stems from the system’s ability to decompose intricate problems into manageable sub-tasks, explore a vast design space with automated efficiency, and learn from both successes and failures. Consequently, industries currently burdened by lengthy development cycles and high costs could experience substantial gains in productivity, innovation, and the creation of optimized, high-performance products. The framework doesn’t merely offer incremental improvements; it proposes a fundamental shift towards automated engineering workflows capable of tackling previously intractable challenges.

The advent of history-aware, intelligently planned engineering workflows represents a fundamental shift in how complex systems are designed and optimized. This work introduces a new paradigm wherein engineering intelligence doesn’t simply solve problems, but learns from the entire history of solutions-successful and unsuccessful-to proactively guide the design process. By integrating past knowledge with forward-looking planning algorithms, the framework achieves a level of automation previously unattainable, significantly reducing development time and fostering genuinely innovative outcomes. This approach promises to reshape engineering practice, moving beyond iterative refinement toward a future where systems are designed with an inherent understanding of what has come before, and a clear vision of optimal future performance.

This agentic workflow leverages both expert knowledge and past training data to intelligently recommend hyperparameter adjustments for reinforcement learning.

The pursuit of Agentic Engineering Intelligence, as detailed in this work, echoes a fundamental truth about complex systems. It isn’t about imposing order, but about cultivating a responsive ecosystem. The framework’s emphasis on memory-augmented reasoning and constraint handling isn’t simply about solving problems; it’s about creating a system that learns to navigate them. As John McCarthy observed, “Every architectural choice is a prophecy of future failure.” This isn’t a pessimistic view, but a call for humility. AEI, by embracing iterative refinement and engineer supervision, acknowledges that perfect design is an illusion, and prepares for the inevitable evolution of the system itself. The focus shifts from static blueprints to dynamic adaptation, a recognition that growth, not control, is the true measure of success.

What Lies Ahead?

This exploration of Agentic Engineering Intelligence, framed as sequential decision processes, feels less like construction and more like tending a garden. The framework attempts to formalize engineering workflows, but it inadvertently highlights how little true formalization is possible. Scalability is, after all, just the word used to justify complexity. The elegance of a constraint-handling system will inevitably be tested by the unforeseen-by the edge cases that always proliferate beyond initial projections.

The promise of memory-augmented reasoning is particularly fraught. The past, so carefully curated within an agent’s knowledge base, is never a perfect predictor. Every optimization will someday lose flexibility, and the very act of remembering risks ossifying the process, stifling genuine innovation. The pursuit of ‘intelligent’ workflows must acknowledge this inherent tension-the trade-off between efficiency and adaptability.

The perfect architecture is a myth to keep engineers sane. Future work should focus not on perfecting the agent, but on cultivating the ecosystem around it-on fostering human oversight, on building in mechanisms for graceful degradation, and, most importantly, on accepting that failure is not a bug, but a fundamental characteristic of complex systems. The real challenge lies not in automating engineering, but in augmenting it with tools that embrace imperfection.

Original article: https://arxiv.org/pdf/2604.07784.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/