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

Hadiza Umar Yusuf; Khouloud Gaaloul

cs.SESep 29, 2026

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

Authors: Hadiza Umar Yusuf, Khouloud Gaaloul

Organizations: University of Michigan-Dearborn Dearborn, USA.

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.

Figures & tables

Fig. 1: Illustrative PMSM field-oriented control (FOC) example: a learned policy block replaces the analytic control law while the surrounding structure stays unchanged.

ID	Categories	Block Types
C1	Continuous Dynamics	Derivative, Integrator, PID Controller, State-Space, Transfer Fcn, Transport Delay, Zero-Pole, FOH, etc.
C2	Discrete Dynamics & State	Delay, Difference, Discrete Derivative, Discrete PID Controller, Discrete State-Space, Memory, ZOH, etc.
C3	Discontinuities & Nonlinearities	Coulomb and Viscous Friction, Dead Zone, Hit Crossing, Rate Limiter, Relay, Saturation, Backlash, etc.
C4	Logic, Conditions, & Events	Combinatorial Logic, Compare To Zero, Logical and Relational Operator, Shift Arithmetic, Extract Bits, etc.
C5	Math & Signal Operations	Abs, Algebraic Constraint, Sum, Bias, Divide, Dot Product, Gain, Math Function, Reshape, Sign, etc.
C6	Ports, Interfaces & Subsystems	Enable, Subsystem, If, If Action Subsystem, Inport, Outport, Switch Case, Trigger, In/Out Bus Element, etc.

TABLE I: Simulink Block Categories and Associated Block Types [ 28 ] . The complete block types are available at [ 27 ] .

ID	Level Type	Description
L1	Core Control Dynamics	Implements the primary control law that generates control actions from system inputs and states.
L2	Signal Constraint Enforcement	Applies constraints or limits to control signals to ensure safe, stable, and valid operation of the system.
L3	Event/Mode Switching Logic	Enables discrete changes in control behavior based on events, thresholds, or operating conditions.
L4	Signal Preprocessing & Computation	Performs transformations or computations to ensure signal compatibility within the control architecture.
L5	Supervisory Coordination	Manages coordination between multiple control elements or modes, ensuring proper sequencing and integration.
L6	Estimation / Observer Support	Provides state estimation or variable reconstruction to supply information not directly measured.

TABLE II: Functional Role Levels (L1–L9) Derived for Controller Characterization. Each level describes a distinct functional purpose within Simulink-based CPS architectures.

Fig. 3: Participants’ experience of controller designs

ID	Model Name	Type	Size	ID	Model Name	Type	Size
M01	Abstract Fuel Control	DRL	400	M07	PMSM Control	PID	2601
M02	Neural Network	DNN	699	M08	Stochastic Fault Tolerance	SBC	273
M03	Temperature Control	FLC	107	M09	Adaptive Cruise Control	MPC	361
M04	Artificial Pancreas Control	FLC	276	M10	Missile Guidance System	PID	396
M05	Steam Condenser	DNN	172	M11	Rolling Mill	LQR	211
M06	Nonlinear Guidance	CLC	355	M12	House Heating System	MPC	200

TABLE III: 12 representative Simulink CPS models used in the survey. Size denotes controller subsystem block count.

Cat	BC(T)	$\mathit{Cov_{c}}$ (T)	BC(AI)	$\mathit{Cov_{c}}$ (AI)	FRL	BC(T)	$\mathit{Cov_{c}}$ (T)	BC(AI)	$\mathit{Cov_{c}}$ (AI)
C1	8	25.8%	0	0.0%	L1	12	38.7%	10	35.7%
C2	7	22.6%	5	17.9%	L2	5	16.1%	0	0.0%
C3	6	19.4%	0	0.0%	L3	2	6.5%	2	7.1%
C4	0	0.0%	1	3.6%	L4	5	16.1%	8	28.6%
C5	2	6.5%	1	3.6%	L5	1	3.2%	2	7.1%
C6	3	9.7%	3	10.7%	L6	0	0.0%	0	0.0%

TABLE IV: Literature-derived block counts (BC) and coverage

\mathit{Cov_{c}}

for taxonomy categories (C1–C10) and functional-role levels (L1–L9) in traditional (T) and AI-enabled (AI) controller descriptions.

	Coverage ( $\mathit{Cov_{c}}$ , %)										Mean normalized presence ( $\mathit{MNP_{c}}$ , %)										Model-level prevalence ( $\mathit{MLP_{c}}$ , %)
Cat.	Traditional					AI-enabled			All		Traditional					AI-enabled			All		Traditional					AI-enabled			All
	PID	MPC	LQR	SBC	CLC	DNN	DRL	FLC	T	AI	PID	MPC	LQR	SBC	CLC	DNN	DRL	FLC	T	AI	PID	MPC	LQR	SBC	CLC	DNN	DRL	FLC	T	AI
C1	0.3	0.0	13.5	0.0	0.0	0.7	0.0	0.6	0.4	0.3	5.3	0.0	16.8	0.0	0.0	1.7	0.0	0.4	3.9	0.2	45.5	0.0	75.0	0.0	0.0	33.3	0.0	20.0	25.8	6.5
C2	0.7	0.7	0.0	3.1	0.3	0.3	5.5	0.6	0.7	3.5	1.5	0.5	0.0	1.3	0.3	1.2	5.5	0.3	0.8	4.2	63.6	33.3	0.0	33.3	28.6	33.3	100	20.0	38.7	80.6
C3	0.6	0.0	5.8	0.8	0.0	0.2	0.0	1.2	0.5	0.2	1.4	0.0	3.6	0.3	0.0	0.4	0.0	1.2	0.9	0.2	72.7	0.0	25.0	33.3	0.0	66.7	0.0	40.0	32.3	12.9
C4	1.1	0.0	0.0	4.6	12.5	0.0	1.4	0.3	2.5	0.9	1.5	0.0	0.0	1.9	10.5	0.0	1.4	0.2	3.1	1.1	36.4	0.0	0.0	33.3	100	0.0	100	20.0	38.7	77.4

TABLE V: Category-level structural comparison across controller paradigms. Controller subsystems across traditional families (PID, MPC, LQR, SBC, CLC), AI-enabled families (DNN, DRL, FLC), and aggregated Traditional (T) vs. AI.

Fig. 4: Survey results for controller design selection across benchmark models (n = 13 participants).

	Traditional					AI-enabled			All
Cat.	PID	MPC	LQR	SBC	CLC	DNN	DRL	FLC	T	AI
C1	1.1/3.5	0.4/2.1	1.8/4.7	1.3/3.7	0.2/1.6	0.9/2.7	0.3/1.9	0.3/1.6	1.0/3.1	0.5/2.1
C2,C4	1.0/3.3	1.0/3.5	0.3/2.2	1.5/3.9	1.2/3.9	0.9/2.9	1.9/4.8	0.9/3.1	1.0/3.3	1.2/3.6
C3	0.9/3.1	0.4/2.3	0.3/1.9	1.2/3.5	0.2/1.9	0.7/2.8	1.0/3.2	1.0/3.3	0.6/2.5	0.9/3.1
C8	0.5/2.3	1.8/4.9	1.3/4.0	0.6/2.6	0.9/3.3	1.1/3.3	1.6/4.5	1.6/4.6	1.0/3.4	1.4/4.1
C7,C9	1.6/4.1	1.8/4.5	1.2/4.2	1.4/4.2	1.6/4.2	1.8/4.7	1.5/4.3	1.5/4.2	1.5/4.2	1.6/4.4

TABLE VI: Survey comparison of taxonomy categories: average presence (0–2) / average essence (1–5) per family (Traditional: PID, MPC, LQR, SBC, CLC; AI-enabled: DNN, DRL, FLC) and aggregated (T vs. AI).

Explore similar work

Structured Representation Learning for Behavior Cloning: How can we learn to safely control a nuclear power plant?
Oct 5, 2026Perceval Beja-Battais, Alain Grosset{ê}te, Nicolas VayatisLearning-Based Control
CB
Position: AI Safety Requires Effective Controllability
May 26, 2026Yige Li, Yunhao Feng, Jun SunArtificial Intelligence Safety Controllability
Singapore Management University · Ant Group
SMSI: System Model Security Inference: Automated Threat Modeling for Cyber-Physical Systems
Apr 26, 2026RoÝah Radaideh, Ali KhreisCyber-Physical System Neuro-Symbolic Framework
School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, ON, Canada

Oct 5, 2026, cs.LGJ/K move · Enter open · S save

Structured Representation Learning for Behavior Cloning: How can we learn to safely control a nuclear power plant?

Perceval Beja-Battais, Alain Grosset{ê}te, Nicolas Vayatis

Learning-Based Control

Open paper AskPDF arXiv

Learned models for industrial control are usually judged by aggregate accuracy, but accuracy at the component level does not guarantee safety once it is embedded in the system it is meant to serve. We study this gap on a behavior-cloning task: imitating an expert Nonlinear Model Predictive Control (NMPC) policy for load-following of a Pressurized Water Reactor (PWR), an industrial system with tight safety constraints. We propose a structured architecture encoding variables from each timescale into separate latent spaces, reflecting the physical decomposition of the system, before training a controller to imitate the expert on the product latent space. On long-horizon rollouts, separated embeddings improve both accuracy and feasibility compared with a shared-embedding baseline. Sensitivity analysis further shows that our model yields interpretable representations aligned with the system's physics. However, standalone deployment still leaves several percent of trajectories infeasible regardless of the architecture. Using our method to warmstart the NMPC optimizer rather than acting standalone, we recover full feasibility and near-optimal cost while still cutting computation time by

\sim

15% relative to the expert controller, and even more for abrupt operating changes.

Abstract

Explore similar work

Structured Representation Learning for Behavior Cloning: How can we learn to safely control a nuclear power plant?

Position: AI Safety Requires Effective Controllability

SMSI: System Model Security Inference: Automated Threat Modeling for Cyber-Physical Systems