AI‑PROPELLER: AI‑Driven Propeller Design & Optimization

• AI‑PROPELLER is a multi-domain concept that integrates AI into propeller design, performance prediction, and fault diagnosis across marine, robotics, and compiler optimization applications.
• In marine engineering, it reduces design complexity by combining inverse modeling, surrogate optimization, and active subspace methods, achieving high prediction accuracy and design efficiency.
• In robotics and compilers, it enables sensor-based state estimation and data-driven code layout optimization, demonstrating robust performance in both physical system control and software engineering.

AI‑PROPELLER is a label used in several distinct research contexts. In marine engineering, it denotes AI‑assisted propeller design systems that couple numerical simulation with inverse modeling, generative modeling, or reduced‑order optimization to produce geometries matching prescribed thrust, power, and efficiency targets. In aerial and underwater robotics, it denotes propeller‑centric sensing, diagnosis, and control pipelines that infer vehicle state or propeller health from wake, vibration, acoustics, or event streams. In a separate compiler context, “AI‑PROPELLER” names a warehouse‑scale interprocedural code layout optimizer built on the Propeller post‑link optimizer (Vardhan et al., 2023, Chen et al., 24 Apr 2026, Ananda et al., 28 May 2026).

1. Scope, antecedents, and common formulation

Across its engineering uses, AI‑PROPELLER refers to systems in which AI is not merely appended to a propeller workflow, but is inserted at a technically central location: inverse design, performance prediction, state estimation, anomaly detection, or fault management. A recurring formulation is to replace or accelerate an iterative loop of the form “geometry $\rightarrow$ simulation $\rightarrow$ redesign” by learning a map from requirements or sensor signals to design or state variables. This suggests a unifying interpretation of AI‑PROPELLER as propeller‑centric computational intelligence, even though the individual implementations differ substantially.

Several precursor lines predate the explicit label. One strand developed a practical online optimization scheme for a DC‑motor–driven variable‑pitch propeller that minimizes electrical power for a commanded thrust by computing the optimum pitch angle online rather than relying on precomputed lookup tables (Cohen et al., 2013). Another established low‑cost experimental techniques for identifying thrust–torque, thrust–PWM, and thrust–angular‑speed relations for quadcopter propellers, with the stated aim of producing a dynamic model suitable for autonomous flight and fail‑safe control even without a reaction torque sensor (Rible et al., 2020). These works did not use the name AI‑PROPELLER, but they established two themes that later reappear: online adaptation and data‑driven propeller modeling.

A second common pattern is the separation between a fast learned component and a physics‑based validator. In the marine design papers, the learned model proposes geometries or reduces the search space, while OpenProp or PropElements remains the authoritative evaluator (Vardhan et al., 2023, Chen et al., 24 Apr 2026). In UAV fault pipelines, learned classifiers operate on signals generated by a hybrid simulator or by controlled experiments, but the underlying vehicle dynamics remain physics‑based (Tong et al., 2023, Arasanipalai et al., 2020). In propeller sensing, neural estimators decode structured but turbulent wake or event signals that would be difficult to invert analytically (Wang et al., 2024, Thakur et al., 20 Apr 2026).

2. Marine propeller design: inverse modeling, surrogate assistance, and reduced search spaces

One explicit use of the term appears in a hybrid Surrogate Assisted Optimization framework for marine propeller design (Vardhan et al., 2023). There the requirement vector is

$r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$

and the design objective is formulated as

$\max_{g \in \mathcal{G}} \eta(r,g),$

with a secondary emphasis on minimizing design time. The motivation is unusual: OpenProp is relatively cheap to evaluate, but the combined requirement–geometry space is estimated to be of order $10^{38}$, so the bottleneck is not per‑evaluation cost but astronomical search complexity. The method therefore moves the surrogate outside the optimization loop and treats design as an inverse problem rather than replacing the simulator directly.

OpenProp itself is used as the numerical core. For a given geometry, it solves a lifting‑line‑based optimization with the auxiliary function

$H = Q + \lambda_1 (T - T_s),$

and stationarity conditions

$$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$

The simulator returns thrust, torque, circulation distribution, and efficiency. On top of this, the SAO framework trains a Random Forest regressor and a fully grown Decision Tree on about $0.205$ million feasible design points. The Random Forest uses $K=100$ trees, while the Decision Tree acts as a one‑to‑many “memory map.” About 90% of held‑out test points have efficiency predicted within 5% relative error, and the resulting surrogate‑seeded GA reaches better solutions than a randomly initialized GA over the same evaluation budget (Vardhan et al., 2023).

A related but distinct reduction strategy appears in the PRELICA propeller campaign, where a BladeX parameterization with as many as 20 shape parameters is analyzed using Active Subspaces (Mola et al., 2019). The quantities of interest are $K_T$, $\rightarrow$0, the tip vortex–induced maximum pressure $\rightarrow$1, and its frequency $\rightarrow$2. The gradient covariance matrix

$\rightarrow$3

is eigendecomposed to reveal that, for the studied PPTC SVA–VP1304 campaign, the main performance indices are, at a first but rather accurate approximation, only depending on a single parameter which is a linear combination of the original geometric ones. Using 1100 PROCAL potential‑flow simulations and an 80/20 split for response‑surface fitting and validation, the work reduces a 20‑dimensional pitch/camber parameter space to an essentially one‑dimensional optimization coordinate (Mola et al., 2019).

Taken together, these results define a marine AI‑PROPELLER pattern in which search is compressed before high‑fidelity optimization begins: either by learning a direct requirement‑to‑design inverse map or by projecting the design space onto a low‑dimensional active coordinate. This suggests that, in marine propeller design, AI‑PROPELLER is less about replacing hydrodynamics than about structuring the combinatorics of design exploration.

3. Generative performance‑to‑design synthesis

A more explicitly generative version of AI‑PROPELLER is presented in a modular performance‑to‑design framework for marine propellers (Chen et al., 24 Apr 2026). Each propeller is encoded as a 162‑dimensional vector comprising six radial distributions—chord length, skew, maximum thickness, rake, pitch, and maximum camber—sampled at 27 normalized radial stations. A physics‑based data synthesis pipeline built with PropElements generates a database of over 20,000 four‑ and five‑bladed geometries and 278,760 $\rightarrow$4 pairs for surrogate training. The conditioning vector for inverse design is

$\rightarrow$5

Two conditional generative models are studied: a $\rightarrow$6-cVAE and a latent diffusion model. The cVAE is trained with

$\rightarrow$7

while the latent diffusion model is trained in a 64‑dimensional VAE latent space. A separate MLP surrogate maps

$\rightarrow$8

with test performance $\rightarrow$9 for $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$0, $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$1 for $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$2, and $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$3 for $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$4. On PropElements validation, the cVAE reaches mean errors of about $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$5 in $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$6 and $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$7 in $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$8, while the latent diffusion model is less accurate but substantially more diverse: its spread coefficient rises to $r = (\text{thrust},\ \text{velShip},\ \text{RPM}),$9 for 4‑blade and $\max_{g \in \mathcal{G}} \eta(r,g),$0 for 5‑blade propellers, compared with about $\max_{g \in \mathcal{G}} \eta(r,g),$1 for the cVAE (Chen et al., 24 Apr 2026). The downstream refinement stage uses CMA‑ES under thrust, power, blade‑area ratio, and thickness constraints.

A second generative marine design formulation uses conditional flow matching rather than diffusion or variational inference (Kruger et al., 29 Jan 2026). Here the design vector is

$\max_{g \in \mathcal{G}} \eta(r,g),$2

and the performance labels are

$\max_{g \in \mathcal{G}} \eta(r,g),$3

The model learns $\max_{g \in \mathcal{G}} \eta(r,g),$4 by regressing a neural vector field in the CFM loss

$\max_{g \in \mathcal{G}} \eta(r,g),$5

Training data are produced by 3000 OpenProp simulations, split into 2000 training and 1000 test examples. The paper also introduces pseudo‑label augmentation: forward surrogate models $\max_{g \in \mathcal{G}} \eta(r,g),$6 predict $\max_{g \in \mathcal{G}} \eta(r,g),$7 for additional Latin‑hypercube‑sampled geometries, and the resulting pseudo‑labeled sets can improve generative accuracy, especially for the more difficult $\max_{g \in \mathcal{G}} \eta(r,g),$8 target (Kruger et al., 29 Jan 2026).

These generative models formalize a central inverse‑design fact already visible in earlier surrogate papers: the mapping from performance targets to geometry is one‑to‑many. AI‑PROPELLER in this sense denotes not a single “optimal propeller,” but a learned conditional distribution over feasible propellers.

4. Fault diagnosis, degradation assessment, and fault‑tolerant control

In UAV and UAM research, AI‑PROPELLER frequently denotes diagnosis or resilience rather than design. One acoustic strand builds the Acoustic Dataset for Crack of Drone Propeller (ADCP) for non‑destructive detection of ripped and broken drone propellers under microphone‑angle and throttle variations (Lee et al., 2 Mar 2025). The dataset contains 4,320 normal samples and 1,440 abnormal samples, recorded at 48 kHz with microphone angles $\max_{g \in \mathcal{G}} \eta(r,g),$9 and throttle settings $10^{38}$0. The anomaly detector is a fully connected autoencoder trained on FFT/STFT features from normal data only. Per‑angle models achieve their best performance around $10^{38}$1 and $10^{38}$2, and a global model trained on all conditions yields F1 scores of about $10^{38}$3 for ripped and $10^{38}$4 for broken, with ROC AUC values about $10^{38}$5 and $10^{38}$6, respectively (Lee et al., 2 Mar 2025).

A second strand uses a hybrid data generation model for quadrotor propeller fault localization and damage classification (Tong et al., 2023). Here PX4 provides target RPMs, LSTM propeller models map RPM to thrust and torque under normal, bent, and cracked conditions, and a rigid‑body quadrotor model propagates the state. The propeller‑level LSTM regressors achieve low average test errors—for example, $10^{38}$7 thrust error for the normal propeller model and $10^{38}$8 thrust error for the cracked model—and the resulting flight data are fed to a CNN classifier over 16 fault classes. Simulation tests report diagnosis accuracy over 80%, and after calibration for real‑vehicle imbalance the sim‑real experiment reaches $10^{38}$9 accuracy, with precision $H = Q + \lambda_1 (T - T_s),$0 and recall $H = Q + \lambda_1 (T - T_s),$1 on the restricted real‑flight label set (Tong et al., 2023).

For active fault isolation in multirotors, a model‑based vibration approach deliberately perturbs control inputs to isolate blade damage using only IMU data (Baldini et al., 8 Apr 2026). On an octarotor, the key oscillatory force term in hover is

$H = Q + \lambda_1 (T - T_s),$2

which places the dominant fault signature in the rotor‑speed band. The method modifies a QP control allocation so that one rotor at a time is driven near a desired $H = Q + \lambda_1 (T - T_s),$3, creating a clean frequency window for fault isolation. The study evaluates 9600 simulations and finds that isolation errors in the confusion matrices are essentially false negatives rather than wrong‑rotor assignments, but performance depends strongly on structural damping (Baldini et al., 8 Apr 2026).

A more control‑centric use of AI‑PROPELLER appears in reinforcement‑learning‑based recovery from mid‑flight propeller loss (Arasanipalai et al., 2020). Separate PPO controllers are trained for 4, 3, and 2 opposing functional propellers, and LSTM fault detectors trigger mode switching. Average detection delay is about $H = Q + \lambda_1 (T - T_s),$4 s for the first propeller loss and $H = Q + \lambda_1 (T - T_s),$5 s for the second. Under the learned 3‑ and 2‑propeller policies, the vehicle settles into spinning equilibria with much smaller yaw rates than the analytic baseline used for comparison, while still tracking waypoints (Arasanipalai et al., 2020).

5. Propeller‑centric sensing and relative‑state estimation

A different family of AI‑PROPELLER systems treats propeller‑induced flow or vision signatures as a sensing modality. In underwater robotics, an artificial lateral line with three pressure sensors can infer the lateral state of a leader propeller from wake measurements (Wang et al., 2024). The input to the estimator is a sliding window

$H = Q + \lambda_1 (T - T_s),$6

with $H = Q + \lambda_1 (T - T_s),$7 sensors and a typical sequence length $H = Q + \lambda_1 (T - T_s),$8. A 1D CNN plus BiLSTM extracts spatiotemporal features, and three task heads jointly estimate lateral displacement $H = Q + \lambda_1 (T - T_s),$9, speed class $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$0, and direction $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$1. With task weights tuned by Whale Optimization Algorithm, the system reaches normalized RMSE $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$2 and speed/direction accuracies $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$3 and $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$4 at $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$5 mm, with somewhat degraded but still strong performance at $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$6 mm (Wang et al., 2024).

For aerial multi‑UAV estimation, event‑based propeller sensing extracts per‑propeller frequencies from 10 ms event chunks and feeds them into kinematic filters (Thakur et al., 20 Apr 2026). The thrust model is

$$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$7

and a second filter estimates orientation from ellipse geometry recovered from a spinning propeller disc. On five outdoor sequences, propeller frequency estimation reaches a mean absolute percentage error of about $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$8, while relative state estimation achieves mean RMSE values $$\frac{\partial H}{\partial \Gamma(i)} = 0,\qquad \frac{\partial H}{\partial \lambda_1} = 0.$$9 m, $0.205$0 m, and $0.205$1 m on position axes, and $0.205$2 and $0.205$3 on roll and pitch (Thakur et al., 20 Apr 2026). This establishes propeller frequency as a passive visual surrogate for thrust input in decentralized relative localization.

Event cameras have also been used for direct propeller detection. EVPropNet trains on simulated event frames generated from a geometric propeller model and transfers to real event data without fine‑tuning (Sanket et al., 2021). The network has about 2.7M parameters and can run at up to 35 Hz on a 2 W power budget via EdgeTPU deployment. It reports an overall detection rate of 85.1% even when 60% of the propeller is occluded, and the resulting detections support real‑world tracking and mid‑air landing experiments with success rates of 92% and 90%, respectively (Sanket et al., 2021).

These sensing systems share a notable methodological stance: they do not attempt full analytical inversion of wake turbulence or blade motion. Instead, they learn compact propeller‑specific observables—pressure patterns, event counts, ellipse geometry, spectral peaks—from which state can be inferred robustly.

6. AI‑PROPELLER in compiler optimization

In a distinct and terminologically separate usage, AI‑PROPELLER is the name of a warehouse‑scale interprocedural code layout optimizer built on the Propeller post‑link optimizer (Ananda et al., 28 May 2026). Here “Propeller” refers not to a physical rotor but to a profile‑guided binary layout system. The paper addresses fine‑grained interprocedural code layout, a problem made difficult by a combinatorially intractable search space and call‑return semantics. In the cited warehouse‑scale example, the application contains about 1.4 million functions and about 40 million basic blocks.

The system extends Propeller through Magellan, an agentic workflow in which AlphaEvolve mutates compiler heuristic code and Vizier tunes exposed numerical hyperparameters. Rather than relying on approximate static cost models, AI‑PROPELLER evaluates multiple layout variants on actual hardware and uses measured performance counters as the reward signal. The evolved policy modifies parameters such as Chain Split Threshold and Forward/Backward Memory Offset, enabling fine‑grained interprocedural placement of basic block sequences while preserving build scalability (Ananda et al., 28 May 2026).

Empirically, the optimizer reports execution‑time improvements of 1.6% on LLVM Clang and 0.23% on a warehouse‑scale Search workload relative to strong FDO and PLO baselines, with p‑values close to zero. An ablation that forces all blocks of a function to remain contiguous reduces the Clang gain from 1.6% to about 0.3%, indicating that most of the improvement arises from interprocedural reordering rather than better intraprocedural tuning (Ananda et al., 28 May 2026).

This compiler usage broadens the encyclopedia sense of AI‑PROPELLER: the term can name either a propeller‑centric AI system in physical engineering or an AI‑evolved extension of the Propeller optimizer in code generation. The shared element is methodological rather than physical—learned or evolved guidance embedded into a previously heuristic optimization loop.