Importance of understanding aerodynamic phenomena and behaviour is paramount in aerospace, automotive, energy or even construction industry. Accurate prediction and analysis of aerodynamic forces and flow patterns are essential for optimizing the performance, efficiency, and safety of products. Traditionally, Computational Fluid Dynamics (CFD) has been the cornerstone for simulating and analyzing aerodynamic behavior. However, CFD utilizes numerical methods that heavily rely on quality mesh, and making a good mesh is of utmost importance, which requires a qualified domain expert. Thus, CFD is computationally expensive and time-consuming, particularly when dealing with complex geometries or parametrized steady-state studies with different flow conditions.

In recent years, the advent of Physics-Informed Neural Networks (PINNs) has opened new avenues for enhancing CFD simulations. PINNs integrate the governing physical laws of fluid dynamics directly into the neural network architecture, and don’t require mesh generation. Meshless computation enables quicker predictions of aerodynamic behavior without need for domain expert.

Siml.ai enables users to create external aerodynamics simulations on any chosen geometry. One trained parametrized model is capable of showing solutions for different chosen flow conditions (flow speed, flow direction) in an inference that is orders of magnitude faster than a full CFD simulation needed for conventional parametric study. It is as simple as importing .stl geometry, specifying boundary conditions and parametrization conditions, training the network on cloud GPU based server, and interpreting results.

Once the model is trained, it can be easily opened in Simulation Studio, where the simulation settings automatically show the parameterized variables which users can set and change between each simulation iteration, which usually takes just few seconds. Here's a quick peek into parameterized simulaton of external aerodynamics of high-rise building done in Simulation Studio and visualized by applying vector field mode and various axis slices:

Key takeaways:

• PINNs provide 100-1000x faster solutions on inference than CFD simulations

• No AI expert needed - focus just on physics

• No mesh needed - Siml.ai will sample point cloud from geometries for training PINN models, you can change how much points are sampled by changing Batch size number in Constraint node settings (higher the batch size, better accuracy of the trained model, but longer traning times and higher requirements on the computational power, possible to require higher Tier selection when creating Environment for a training run)

• PINNs reduce overall time and cost of physics simulations