The Urgent Need to Understand Supercooled Large Droplet Icing

Supercooled large droplet (SLD) icing stands as one of the most persistent hazards in aviation, directly linked to several high-profile accidents including the 1994 ATR-72 crash near Roselawn, Indiana. When an aircraft flies through clouds containing supercooled water droplets—liquid water that remains unfrozen below 0°C—those droplets can instantly freeze on impact with the airframe. Large droplets, defined by the FAA in 14 CFR Part 25 Appendix O as those greater than 50 micrometers in diameter, behave fundamentally differently from smaller cloud droplets. They can splash, bounce, and break apart, leading to ice accretion patterns that traditional icing simulations fail to predict. Improving simulation fidelity is not just an academic exercise; it is a critical step toward certifying aircraft for safe flight in all known icing conditions.

What Makes Supercooled Large Droplet Icing Unique

Definition and Size Regimes

Conventional icing certification (Appendix C) covers droplets up to 50 micrometers. SLD conditions extend that range up to thousands of micrometers. Two sub-categories exist: freezing drizzle (50–500 µm) and freezing rain (>500 µm). The sheer size of these droplets alters their inertia, making them less likely to follow airflow streamlines around wings and more likely to impact surfaces aft of protected areas—regions where ice protection systems may not be present.

Ice Type and Morphology

SLD impacts produce mixed ice or glaze ice with complex shapes. Large droplets often spread on the surface before freezing, creating smooth, clear ice that adheres tenaciously. Unlike rime ice formed by small droplets freezing instantly upon impact, glaze ice runs back along the chord, forming horns and ridges that severely disrupt aerodynamics. Simulating this runback and the resulting 3D ice shapes requires modeling unsteady heat transfer, water film dynamics, and solidification simultaneously.

Physical Phenomena That Defy Simple Simulation

Accurate SLD simulation must account for at least six non-linear physical processes:

  • Droplet Splash and Re-Impingement: When a large droplet hits a surface, it can break into smaller secondary droplets that remain airborne and strike further downstream. These satellite droplets create ice in unexpected locations, including on horizontal stabilizers or engine inlets.
  • Bouncing and Rebound: Droplets may bounce off dry or partially-iced surfaces, carrying away mass and energy. The bounce threshold depends on surface temperature, roughness, and liquid film thickness.
  • Film-Rivulet Flow: Unfrozen water from large droplets can form a thin film or break into rivulets under the influence of aerodynamic shear. Predicting where that water freezes requires solving the coupled Navier-Stokes and energy equations on a moving boundary.
  • Ice Bridging: Ice may accrete at the edges of protected areas, then grow inward to cover the protected surface. This phenomenon can render de-icing boots or electrothermal heaters ineffective if not anticipated in the design.
  • Shadowing and Scavenging: Upstream protuberances (e.g., ice horns) block downstream surfaces from receiving droplets. Meanwhile, some droplets are captured by earlier ice growth, altering the mass distribution.
  • Thermal Inertia and Latent Heat: The release of latent heat of fusion as droplets freeze raises local surface temperatures, potentially keeping water liquid longer. This feedback loop influences where and how fast ice builds.

Key Challenges in Computational Simulation

Complex Fluid Dynamics and Multiscale Phenomena

Modeling SLD icing requires resolving droplet trajectories in a turbulent flow field while also capturing impact dynamics that occur at microsecond scales. Droplets range from 50 µm to several millimeters, meaning the simulation must bridge orders of magnitude in length and time. Eulerian–Lagrangian frameworks are commonly used, but the coupling between the continuous air phase and the dispersed water droplets introduces numerical stiffness. High-fidelity large-eddy simulation (LES) can improve accuracy but at a computational cost that remains prohibitive for routine certification work.

Thermal and Phase-Change Modeling

The freezing process involves unsteady heat conduction into the solid substrate, convective cooling from the airflow, evaporative cooling, and latent heat release. Most icing codes simplify this by assuming a single freezing temperature (0°C) and ignoring subcooling hysteresis. For SLD, however, droplets can remain supercooled to -20°C or below, leading to non-equilibrium freezing rates. Accurately capturing the transition from liquid film to solid ice demands a multi-phase, non-isothermal solver that few tools currently provide.

Droplet Size Distribution and Cloud Variability

Real icing clouds contain a continuous spectrum of droplet sizes. Standard certification tests use a median volume diameter (MVD) and liquid water content (LWC), but SLD clouds can have multiple modes. For example, freezing drizzle may include both small droplets from the drizzle mode and larger droplets from a rain mode. Simulating every droplet size is computationally intractable; instead, researchers use representative bins (e.g., 10–40 bins) and interpolate. However, binning introduces discretization error, especially for the splash and bounce models that are highly nonlinear with diameter.

Surface Interaction and Roughness Effects

Aircraft surfaces are not smooth. Paint roughness, ice roughness from earlier accretion, and erosion change the local flow field and heat transfer. Roughness affects the critical Weber and Reynolds numbers that determine whether a droplet splashes or sticks. Most simulations either assume a smooth surface or a uniform roughness height, neither of which captures the stochastic nature of real ice roughness. Recent research uses fractal roughness models, but these have not yet been integrated into production-grade icing solvers.

Environmental Variability and Uncertainty

SLD conditions involve combined uncertainties in temperature, humidity, pressure, wind vector, and cloud microphysics. Even small changes can shift the ice shape from benign to catastrophic. Probabilistic simulation approaches (e.g., Monte Carlo methods or polynomial chaos) are gaining traction, but they multiply the required number of runs, making computational efficiency a recurring bottleneck.

Current Simulation Approaches and Their Limitations

Computational Fluid Dynamics (CFD) Methods

The industry standard for ice accretion simulation is a quasi-steady approach: first compute the airflow field (often using Reynolds-averaged Navier-Stokes, RANS), then compute droplet impingement via Lagrangian particle tracking or Eulerian scalar transport, finally solve a thermodynamic model (e.g., Messinger model) for the ice growth. Tools like NASA's LEWICE, ONERA's IGLOO, and commercial codes such as FENSAP-ICE and STAR-CCM+ have evolved to include SLD-specific models, including splash and bounce subroutines. However, these models rely on empirical correlations derived from a limited set of wind tunnel tests. Extrapolating to new geometries or flight conditions introduces significant uncertainty.

Experimental Validation as a Check

Laboratory icing tunnels, such as the NASA Icing Research Tunnel at Glenn Research Center, remain the gold standard for validation. Yet even these facilities struggle to produce true SLD clouds—large droplets settle out of the airflow, collide with tunnel walls, and break up before reaching the model. Partial solutions include upscaled models and matching non-dimensional parameters (Weber, Reynolds, freezing fraction), but the extrapolation from tunnel to full-scale flight is not always reliable. Field tests with instrumented aircraft, like the FAA's SLD research campaign with a Twin Otter, provide crucial data but are expensive and weather-dependent.

Empirical vs. Physics-Based Models

Most splash and bounce models in use today are empirical—for example, the relationship between the incoming droplet kinetic energy and the mass fraction that remains on the surface. Physics-based models based on direct numerical simulation (DNS) of droplet impact have been developed for simple cases, but they remain too computationally expensive for full aircraft geometries. A hybrid approach, where DNS-generated data trains lower-order models (e.g., through neural networks), is an active research area.

Regulatory and Certification Realities

Since 2015, the FAA's 14 CFR Part 25 Appendix O and the corresponding EASA CS-25 Appendix O have mandated that transport-category aircraft must demonstrate safe flight in SLD icing conditions. This regulation applies to new type designs and, for some features, to in-service aircraft. Certification can be achieved through a combination of analysis (simulation) and tests (tunnel or flight). The analytical tools must be validated against SLD-specific datasets. As a result, aircraft manufacturers and their suppliers are under pressure to adopt simulation methods that have demonstrable accuracy for both Appendix C and Appendix O conditions. This regulatory driver is accelerating investment in higher-fidelity models and better validation campaigns.

Machine Learning and Reduced-Order Models

Given the computational cost of high-fidelity CFD, machine learning is being explored to replace or accelerate the most expensive sub-models. For example, a convolutional neural network (CNN) can predict the splash mass ratio from images of droplet impact events, bypassing the need for a full multiphase simulation. Similarly, reduced-order models (ROMs) trained on a database of high-fidelity results can approximate ice shapes in milliseconds, enabling probabilistic certification studies or real-time onboard hazard assessment.

Coupled Multi-Physics and High-Fidelity Simulation

Future simulation systems will couple the droplet impact code with a structural solver to predict the effect of ice on control surface effectiveness, flutter margins, and load distribution. This requires transferring aerodynamic and mass-loading data across meshes with different discretizations. Research frameworks like SU2 (an open-source CFD code) now include conjugate heat transfer and are being extended for icing. As exascale computing becomes available, direct simulation of droplet clouds with millions of droplets in a full aircraft configuration may become feasible within a decade.

Real-Time Data Assimilation

In-flight icing detection systems (e.g., probes on the fuselage) can measure LWC, MVD, and temperature in real time. If coupled with an onboard simulation that runs continuously, the aircraft could update its ice protection logic or even predict when severe conditions will occur. The computational challenge is extreme, but advances in edge computing and reduced-order modeling make this a realistic goal for future autonomous or remotely piloted aircraft.

Conclusion

Simulating supercooled large droplet icing is among the most demanding multi-physics problems in aerospace engineering. It couples fluid dynamics at multiple scales, unsteady heat transfer with phase change, and chaotic surface interactions—all under the pressure of regulatory deadlines and safety-critical applications. While significant progress has been made, especially in splash and bounce modeling and in regulatory frameworks, no single simulation tool today captures the full complexity of SLD icing with certainty. The path forward lies in hybrid physics–machine learning approaches, tighter integration with experimental validation, and continued investment in both computational and experimental infrastructure. For the aviation industry, these improvements are not optional; they are essential to ensuring that aircraft can operate safely in the full range of icing conditions that nature can produce.