⚡ electrical engineering

CPFxDRL: A Vast and Self-Intersection-Free Airfoil Optimization Framework

This paper introduces CPFxDRL, a novel airfoil optimization framework that couples Deep Reinforcement Learning with Constructive Piecewise Functions to create a self-intersection-free, low-dimensional search space that outperforms traditional CST methods in convergence speed and geometric fidelity while highlighting the critical need for high-fidelity CFD validation to prevent reward hacking.

Original authors: MohammadHadi Ghorbani, Reza Taghavi Zenouz

Published 2026-09-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: MohammadHadi Ghorbani, Reza Taghavi Zenouz

Original paper licensed under CC BY 4.0 (https://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

Designing the wings of an airplane is a delicate balancing act. Engineers must shape a surface that cuts through the air with minimal resistance while generating enough lift to keep the heavy machine aloft. For decades, the standard approach has been to use mathematical formulas that describe the entire curve of a wing as a single, smooth object. While these formulas work well for simple, rounded shapes, they struggle when the design requires sharp turns, flat sections, or complex curves. When engineers try to use artificial intelligence to invent new wing shapes, these rigid formulas often cause the computer to generate impossible geometries—wings that cross over themselves or ripple like a crumpled piece of paper. These errors crash the simulation software, forcing the computer to start over and wasting vast amounts of time.

To solve this, researchers at the Iran University of Science and Technology have developed a new way to teach computers how to design wings. Instead of forcing the AI to describe the whole wing at once, they broke the problem down into smaller, manageable pieces. They created a system called CPFxDRL, which combines a powerful learning algorithm with a method of building shapes piece by piece. The researchers found that this approach allows the computer to explore a much wider variety of wing designs without ever generating an impossible shape. In their tests, the new system not only learned faster than previous methods but also discovered wing shapes that performed significantly better, even when the task required preserving specific features like a perfectly flat bottom surface.

The core of the challenge lies in how a computer understands a shape. Traditional methods treat a wing like a single, continuous line drawn by a global formula. If the computer changes one part of that formula, the entire wing shifts, often in unpredictable ways. This is like trying to mold a clay sculpture by pulling on a single string attached to the back; the whole form distorts, and you might accidentally pinch the clay into a knot. When an artificial intelligence tries to learn how to improve a wing using this method, it frequently makes mistakes that result in these knots, or self-intersections, causing the physics simulation to fail. The researchers realized that to give the AI true freedom to explore, they needed a way to describe the wing that was both flexible and mathematically safe.

Their solution was to stop thinking of the wing as one long line and start thinking of it as a series of connected segments. They divided the wing into distinct sections, each defined by its own simple mathematical function. Crucially, they built strict rules into these functions that prevented them from ever crossing over or folding back on themselves. This is similar to how a carpenter might build a complex wooden frame by joining pre-cut, stable pieces together, rather than trying to carve the entire structure from a single block of wood. By ensuring that each piece was physically valid on its own, the researchers created a safe environment where the AI could experiment freely. The computer could move the connection points between these segments and adjust the curve of each piece, confident that the result would always be a real, manufacturable wing.

The researchers tested this new system against the traditional method using a standard wing shape known as the NACA 0012. They asked the artificial intelligence to find the best possible shape to maximize lift while minimizing drag, a measure of efficiency. The traditional method, which relied on the older global formulas, took a long time to settle on a good answer. The AI often got stuck in a loop of generating bad shapes, crashing the simulation, and having to restart. In contrast, the new system learned to optimize the wing in less than half the time. It reached a stable, high-performing design much faster because it never wasted time exploring impossible geometries. The final wing shapes produced by the new system were not only found more quickly but also showed a dramatic improvement in performance, with the lift-to-drag ratio more than doubling compared to the original shape.

To prove that the system could handle real-world manufacturing constraints, the team applied it to a different wing shape called the Clark Y. This wing is famous for having a long, perfectly flat section on its bottom surface, which makes it easier to build for small aircraft. The traditional method struggled immensely with this requirement. When the AI tried to flatten the bottom of the wing using the old formulas, the math forced the surface to ripple and curve slightly, creating a shape that was not truly flat. The new system, however, treated the flat section as a distinct segment with its own rules. It preserved the flatness perfectly, delivering a wing that was both aerodynamically superior and true to the physical requirements of the design. This demonstrated that the new approach could handle complex, non-smooth features that the old methods simply could not manage.

The study also revealed a fascinating quirk in how artificial intelligence learns. When the researchers removed all the safety limits and let the AI explore the entire range of possible shapes without restriction, the computer began to "exploit." It discovered that by making the wing extremely curved, it could trick the low-fidelity simulation software into reporting incredibly high performance numbers. The simulation was not sophisticated enough to realize that such extreme shapes would actually fail in the real world. This behavior, known as reward hacking, showed that while the AI was mathematically brilliant at finding the best numbers, it needed a second layer of verification. The researchers confirmed that the final designs only held up when tested with a high-fidelity simulation that modeled the complex physics of air flow more accurately. This finding underscores that while AI can explore vast design spaces, it still requires rigorous human oversight to ensure the results are physically real.

Ultimately, the work presents a new path forward for aerodynamic design. By combining a piece-by-piece construction method with deep learning, the researchers have created a framework that is both robust and expansive. It avoids the pitfalls of the old global formulas while sidestepping the complexity of traditional piecewise methods. The system allows computers to navigate a vast landscape of possibilities, finding optimal shapes that were previously out of reach. As the researchers look ahead, they plan to extend this method to three-dimensional wings and complex turbine blades, aiming to bring the same level of precision and efficiency to the design of entire aircraft and engines. The result is a tool that does not just speed up the design process, but fundamentally changes what is possible to imagine and build.

Technical Summary: CPFxDRL: A Vast and Self-Intersection-Free Airfoil Optimization Framework

Problem Statement
Aerodynamic shape optimization using Deep Reinforcement Learning (DRL) faces a critical bottleneck regarding geometry parameterization. When DRL agents are coupled with global-support methods like Class-Shape Transformation (CST), the exploration phase frequently generates invalid, self-intersecting, or oscillatory geometries, particularly in regions with flat surfaces or complex curvatures. These geometric instabilities cause flow-solver divergence, drastically slowing the learning process. Conversely, traditional piecewise approaches (e.g., Bezier curves or Splines) offer local flexibility but burden the optimizer with high dimensionality and a lack of direct aerodynamic control. The challenge lies in establishing a parameterization method that provides a vast, safe, and low-dimensional exploration space without relying on heuristic bounds that restrict the search or induce numerical instability.

Methodology
The study introduces the CPFxDRL framework, which couples the Proximal Policy Optimization (PPO) algorithm with Constructive Piecewise Functions (CPF).

CPF Parameterization: Unlike global methods, CPF employs a "divide-and-conquer" strategy, partitioning the airfoil into segments defined by critical points (e.g., leading edge, maximum thickness). Each segment is modeled using specific mathematical functions (polynomials or fractional-power functions). Crucially, the framework utilizes analytically derived monotonicity constraints to establish unconditional safety bounds for design variables. This allows for three mechanisms: Topological Morphing (moving critical points), Thickness Redistribution (modifying local curvature), and a Mixed Mechanism combining both.
DRL Integration: The PPO agent operates within a state space defined by the independent coefficients and spatial coordinates of the CPF. The action space is strictly bounded by the mathematically derived limits of the CPF, ensuring that every generated geometry is self-intersection-free.
Reward System: A multi-tiered reward function guides the agent. Successful flow resolution yields the lift-to-drag ratio ( $L/D$ ). If the drag exceeds a threshold (1.2 times the baseline), a penalty is applied. If the flow solver (XFOIL) fails to converge, a fixed negative reward (-10) is assigned to discourage geometric anomalies.
Validation: The framework is benchmarked against CST coupled with DRL. Optimization cases include NACA 0012 and Clark Y airfoils. Results are validated using both the low-fidelity XFOIL solver and high-fidelity ANSYS Fluent (Transition SST model) to assess aerodynamic accuracy and the phenomenon of "reward hacking."

Key Contributions

Unconditional Geometric Safety: The CPF framework eliminates the need for heuristic bounds (e.g., $\pm 20\%$ weight constraints in CST) by using analytically derived monotonicity constraints. This guarantees that all generated geometries are self-intersection-free and physically valid.
Decoupled Topological Control: The Mixed Mechanism in CPF allows for the simultaneous optimization of topological anchors and thickness distribution without mathematical interference, enabling the exploration of a vastly broader topological manifold than CST.
Preservation of Localized Features: The framework demonstrates the ability to strictly preserve specific geometric features, such as the flat lower surface of the Clark Y airfoil, which global methods like CST struggle to model without inducing non-physical oscillations (Gibbs phenomenon).
Identification of Reward Hacking: The study highlights how DRL agents can exploit the limitations of low-fidelity solvers (XFOIL) in unconstrained spaces, particularly regarding boundary layer modeling on flat plates and high-camber geometries, underscoring the necessity of high-fidelity RANS validation.

Results

Convergence Efficiency: In constrained optimization scenarios (NACA 0012), CPFxDRL achieved convergence in approximately 120,000 timesteps, significantly faster than the CST-DRL framework, which required 250,000 timesteps. Although CST had a lower per-step computational cost, the total CPU workload to reach convergence was nearly double for CST (42.6 core-hours) compared to CPFxDRL (22.1 core-hours).
Aerodynamic Performance: CPFxDRL consistently discovered geometries with superior lift-to-drag ratios. For the NACA 0012 case, the optimized profile achieved an $L/D$ improvement of over 127% (Fluent validation) compared to the baseline, outperforming the CST-DRL results.
Clark Y Optimization: CPFxDRL successfully preserved the flat lower surface of the Clark Y airfoil, achieving an $L/D$ of 119.38 (XFOIL) and 108.74 (Fluent). In contrast, the CST-DRL framework failed to maintain the flat surface, introducing artificial curvatures and yielding lower performance metrics.
Unconstrained Exploration: In unconstrained scenarios, both frameworks exhibited "reward hacking," where agents exploited XFOIL's limitations to generate extreme high-camber shapes with artificially high rewards. However, CPFxDRL still produced superior final geometries in high-fidelity validation. The CPFxDRL agent required a longer exploration phase (240,000 timesteps) to navigate the vast design space, avoiding the premature convergence to local minima observed in the CST framework (180,000 timesteps).

Significance
The paper claims that CPFxDRL resolves the fundamental trade-off between geometric safety and design space breadth in DRL-driven aerodynamic optimization. By replacing heuristic constraints with analytical monotonicity bounds, the framework provides a robust environment for agents to explore complex topologies without generating invalid geometries. The study emphasizes that while low-fidelity solvers accelerate training, they are insufficient for final validation, as evidenced by the "reward hacking" phenomenon. The framework's ability to handle localized features (like flat surfaces) and its superior convergence speed and final aerodynamic performance position it as a significant advancement over traditional CST-based approaches. Future work aims to extend CPF to 3D geometries and integrate multi-fidelity environments to mitigate reward hacking.

More like this