An Empirical Study of Architectural Shift from Traditional to AI-Enabled Simulink Controllers

arXiv:2609.38504 · cs.SE, cs.AI, cs.SY, eess.SY · Submitted 2026-09-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "An Empirical Study of Architectural Shift from Traditional to AI-Enabled Simulink Controllers".

Jane: Effective AI adoption in cyber-physical systems (CPS) depends on embedding design knowledge into engineering practice,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So, looking at this paper's title, "An Empirical Study of Architectural Shift from Traditional to AI-Enabled Simulink Controllers," it immediately tells us they’re doing an empirical look at the structural changes happening when we switch control laws from traditional methods to those powered by AI within the Simulink environment.

Jane: That means they aren't just guessing what happens; they are actually running experiments on sixty-two real-world Simulink models, which gives them a solid foundation for their analysis of these architectural shifts. It sounds like the authors are trying to quantify these changes in a way that moves beyond just reading literature.

Lu: The authors are using a literature-derived taxonomy, which is their method for organizing all those structural categories and functional roles they pulled from existing research, so they can compare traditional and AI-enabled designs using this consistent framework. It’s about imposing order on a very messy field of control system design.

Meng: I'm wondering how much weight the authors put on the practitioner survey part; having the models and the literature is one thing, but getting input from people who actually build these things gives you a really practical sense of what’s happening on the ground.

Lalam: That practitioner involvement suggests they are aiming for something very useful, moving this knowledge out of just academic theory and into something that engineers can actually use when designing their next system.

The paper's summary: Tom: To summarize what the paper is about, it’s essentially an empirical study that uses a specific taxonomy to look at how controller architectures differ or stay the same when moving between traditional control and AI-enabled control paradigms. They analyzed sixty-two Simulink models covering eight different controller types and ten application domains.

Jane: That means they found that there are three main architectural tensions they identified after comparing what the literature says, what the real models show, and what practitioners think is happening in practice. It’s a comparison across three different viewpoints to get a clearer picture of the situation.

Lu: The paper points out that structural scaffolding, specifically subsystem organization, dominates almost every controller structure they looked at regardless of whether it's traditional or AI-enabled, taking up sixty-eight to seventy-two percent of the total controller footprint.

Meng: That dominance is a big finding because it suggests that no matter if you use a classic control method or an AI method, the fundamental way you organize the system—the scaffolding—remains largely the same in terms of physical space consumed.

Lalam: And they also found that while AI-enabled controllers lean heavily on discrete dynamics and user-defined abstraction, which isn't often seen in the literature describing those AI systems, it points to a gap between what we read and what is actually implemented.

The paper's improvements: Tom: The paper suggests a few key improvements based on their findings, focusing on how we can make these AI-enabled systems more robust and easier to understand by shifting from implicit logic to explicit structure.

Jane: They suggest incorporating a "Scaffolding-Aware" design philosophy for the AI components, which means clearly separating the learned policy from all that supporting structural organization like subsystems and data routing. This should help engineers focus on optimizing the algorithm rather than getting bogged down in debugging complex structural interactions.

Lu: Beyond that, they argue we need to enforce explicit constraint enforcement blocks within AI architectures instead of relying only on safety measures implemented during training or reward shaping, which they found to be less reliable for deployment.

Meng: That makes sense from a practical standpoint; if you want high assurance in a real-world setting, you need those constraints visible and enforced at the structural level during deployment, not just tucked away in the training process.

Lalam: The final improvement they push for is developing paradigm-aware verification and traceability mechanisms so we can clearly link the learned control policy back to the specific structural components that support it, which should allow for targeted verification of both performance and structure.

Conclusion: Tom: So, to wrap things up on "An Empirical Study of Architectural Shift from Traditional to AI-Enabled Simulink Controllers," the core implication is that while core control dynamics are shared between traditional and AI methods, the structural approach changes significantly in how they are built.

Jane: Exactly; the shift is localized structurally, with scaffolding being dominant everywhere, but constraint enforcement becoming less visible in deployed AI models compared to what we might expect. This study gives us a baseline for comparing these two approaches empirically across literature, models, and practitioner views.

Lu: For the big picture, this work establishes that understanding the structural organization is as important as understanding the control law itself when evaluating AI-enabled CPS designs. It sets an empirical benchmark for how we should categorize these different design philosophies moving forward.

Meng: From my side, it suggests that tools used for verification and safety assurance need to be built with this paradigm awareness in mind, specifically targeting the explicit scaffolding and constraint issues they highlighted.

Lalam: I feel that this work really helps push us toward making AI-enabled CPS a more transparent process by demanding that we treat the structure of the control system as a first-class element in our design and verification efforts.

Hadiza Umar Yusuf, Khouloud Gaaloul

University of Michigan-Dearborn

cs.SE, cs.AI, cs.SY, eess.SY

Submitted: 2026-09-29

Updated: 2026-09-29

Comments: Accepted at the 33rd Asia-Pacific Software Engineering Conference (APSEC 2026)

Code: https://github.com/AISE-CPS-Research-Lab/CPS-Characterization

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: Effective AI adoption in cyber-physical systems (CPS) depends on embedding design knowledge into engineering practice, and this study addresses the gap in understanding how controller architectures

Key concepts

Taxonomy
A standardized vocabulary of ten structural categories and nine functional roles derived from literature. This system was used to consistently classify different Simulink blocks, allowing researchers to compare traditional and AI-enabled controllers using a unified framework, independent of the control method.
Structural Scaffolding
The finding that subsystem organization (C6) is the most dominant feature in all controller structures across both paradigms. This means that the basic way components are organized provides 68–72% of a controller's structure, suggesting a universal organizational pattern regardless of whether the control law is traditional or AI-based.
Architectural Tensions
Specific structural differences observed between traditional and AI controllers. These include AI models favoring discrete dynamics (C2) and user-defined abstraction (C8), while lacking constraint enforcement blocks (C3). These tensions reveal where the design philosophy shifts between the two paradigms.

Terminology

Summary

Effective AI adoption in cyber-physical systems (CPS) depends on embedding design knowledge into engineering practice, and this study addresses the gap in understanding how controller architectures differ or remain similar across traditional and AI-enabled paradigms. The study analyzes 62 real-world Simulink models spanning 8 controller types and 10 application domains, guided by a literature-derived taxonomy of ten structural categories and nine functional roles to identify three key architectural tensions.

Research Objectives

The empirical evaluation is structured around three research questions designed to compare controller composition across literature, model-based systems, and practitioner perception. The first question examines the shift in how the literature portrays controller structure between traditional and AI-enabled paradigms, focusing on which categories and functional roles receive emphasis or omission in each. The second question investigates what taxonomy categories are most prevalent and structurally dominant in real Simulink controller subsystems, comparing these empirical patterns against the literature-derived expectations from the first question. Finally, the third question seeks to corroborate these structural differences through manual assessment by software engineering practitioners to determine if they recognize these patterns and what implications they carry for software engineering practice.

Taxonomy Construction

The study is guided by a literature-derived taxonomy of ten structural categories (C1–C10) and nine functional roles (L1–L9), providing a consistent vocabulary independent of paradigm. This taxonomy was constructed through three steps: first, literature search and selection, which curated 21 sources covering traditional control (PID, LQR, MPC) and AI-enabled control (DRL, DNN); second, literature extraction of structural categories and functional roles by mapping Simulink block types to these categories; and third, mapping structural categories to functional roles using block-level literature descriptions. The resulting taxonomy groups blocks into ten structural categories (e.g., C1: Continuous Dynamics; C10: Artificial Intelligence RL Agent) and nine functional roles (e.g., L1: Core Control Dynamics Implements the primary control law).

Empirical Dataset and Metrics

The research analyzes 62 real-world Simulink models spanning 8 controller types and 10 application domains, which are classified as traditional or AI-enabled based on whether their control law is realized through a learned component. To quantify structural composition, three metrics are employed: Coverage (Covc), which measures dataset-level structural emphasis; Mean Normalized Presence (MNPc), which measures average within-model prominence by computing each category’s relative share per model; and Model-Level Prevalence (MLPc), which captures how broadly a category is adopted across the dataset. These metrics allow for a comparison of Coverage, MNPc, and MLPc to characterize structural space, within-model prominence, and cross-model adoption.

Key Findings on Architectural Tensions

The empirical evaluation revealed three architectural tensions:

  1. Structural scaffolding dominates all controller structures regardless of paradigm: subsystem organization (C6), dominates all controller structures regardless of paradigm, occupying 68–72% of controller footprint. This highlights that the most striking finding is the sheer dominance of what we term architectural 'scaffolding' across all controller paradigms.

  2. AI-enabled controllers rely heavily on discrete dynamics and user-defined abstraction: AI-enabled controllers rely heavily on discrete dynamics (C2) and user-defined abstraction (C8) despite these being nearly absent in AI literature, exposing a gap between described and implemented architectures.

  3. Constraint enforcement blocks largely disappear from AI-enabled models: constraint enforcement blocks (C3) largely disappear from AI-enabled models even as practitioners expect them, revealing a misalignment between training-time safety and deploy-time structure.

Practice Corroboration and Implications

The survey of 13 practitioners corroborated the structural differences, showing that C7 and C9 are highly dominant in both paradigms (presence: 1.5/1.6), while C8 and C2,C4 are moderately dominant, with AI-enabled controllers showing higher dominance in these categories. The main misalignment identified was regarding C3, where practitioners rated constraint-related structure as more essential than its actual prevalence supports, likely confusing signal-shaping or transformation blocks for constraint mechanisms. These findings suggest that tools should represent scaffolding as a reusable, first-class abstraction and that safety-relevant constraint logic in AI-enabled controllers moves to training-time mechanisms invisible during manual review, necessitating explicit artifacts for inspection.

Conclusion

The study establishes the first empirical baseline for cross-paradigm architecture comparison, demonstrating that while core control dynamics are shared, the shift from traditional to AI-enabled design is structurally localized, with scaffolding dominating and constraint enforcement becoming implicit in AI models. This motivates paradigm-aware tool support for verification and safety assurance in AI-enabled CPS.

Improvements for AI systems

Based on the empirical findings and architectural insights presented in this paper, here are specific improvements that can be made to AI systems integrated into cyber-physical systems (CPS), and what these improved systems could achieve:

The core improvement centers on shifting from implicit, opaque control logic to explicit, traceable architectures. The suggested improvements fall into three main categories: Architectural Redesign, Constraint Visibility, and Verification Strategy.

  1. Incorporate a Scaffolding-Aware Design Philosophy for AI Components

  2. Enforce Explicit Constraint Enforcement Blocks in AI Architectures

  3. Implement Paradigm-Aware Verification and Traceability Mechanisms

The improved AI system can perform the following specific actions:

  1. Generate a more robust and maintainable control law by clearly separating the learned policy (the AI core) from the supporting architectural scaffolding (subsystems, data routing, interfaces). This allows engineers to focus their efforts on optimizing the algorithm rather than debugging complex structural interactions.

  2. Ensure that safety and stability requirements are explicitly enforced during both training and deployment phases by treating constraint blocks as first-class structural elements in the model, rather than relying solely on implicit artifacts from training-time reward shaping or environment design.

  3. Achieve higher assurance levels for safety-critical applications by establishing a clear, verifiable link (traceability) between the learned control policy and the specific structural components (subsystems, data flow) that support it. This allows for targeted verification of both the algorithm's performance and the system's structural integrity.

In summary, these improvements move AI-enabled CPS from being a black-box behavioral shift to a transparent architectural shift, enabling reliable deployment in safety-critical domains by making the structure of the control system as important as its learned behavior.

Abstract

Effective AI adoption in cyber-physical systems (CPS) depends on embedding design knowledge into engineering practice. Yet as AI-enabled components increasingly replace analytically derived control laws, this occurs without a systematic understanding of how controller architectures differ or remain similar across paradigms. We address this gap with an empirical study of traditional and AI-enabled Simulink controllers, guided by a literature-derived taxonomy of ten structural categories and nine functional roles. The study analyzes 62 real-world models spanning 8 controller types and 10 application domains, and surveys 13 practitioners, identifying three architectural tensions. First, subsystem organization dominates all controller structures regardless of paradigm, occupying 68-72% of controller footprint, while core control logic occupies minimal space. Second, AI-enabled controllers rely heavily on discrete dynamics and user-defined abstraction, categories largely absent from AI literature, exposing a gap between described and implemented architectures. Third, constraint enforcement blocks largely disappear from AI-enabled models despite practitioner expectations. This reveals a misalignment where safety mechanisms shift from explicit structure to implicit training-time artifacts, breaking traceability.

Sources

Related papers