This is my attempt to prototype a mobile game idea that uses magnetic forces and fluid simulation inspired by ferrofluids.
I chose Unreal Engine 5 for the task. After setting up the project repo using Perforce and preparing the build pipeline for iOS, I was ready to make the prototype.
I started with the magnetic field, the idea here is when you click on an astroid, it becomes magnetic and attracts the particles. It’s not the best gameplay mechanic I know, but it was a good enough start to test the magnetic forces.
Upon my initial investigation, I realized that although UE5 has a powerful fluid simulation system, it doesn’t work in macOS or iOS. Honestly, I didn’t expect it either, as I’m already surprised that UE5 runs smoothly on M-series MacBooks!
I stumbled upon this brilliant post while searching for another solution. The idea is to keep the number of particles limited (64, for example) and use a rendering technique called ray marching to smooth these particles together into a fluid-like substance. These smoothed spheres are also called metaballs.
As far as I know, Unreal Engine doesn’t support metaballs or even ray marching by default. In the post mentioned above, the author uses a plugin, but it is outdated now. I tried finding other add-ons but didn’t find anything good. Therefore, I decided to create it from scratch. This way, I had more control over the results and the opportunity to learn ray marching along the way.
However, this decision meant that I had to write some HLSL in UE5. I’ve had prior experience writing HLSL code in Unity, and while Unity supports HLSL really well, it’s not the same in Unreal Engine; You can write custom HLSL code, using a Custom node in material editor, but it’s really user-unfreindly (or unuser-friendly?) and you can’t even reference a file and have to use a plain text field to write your code. I was searching for resources on that and found this epic (pun intended) live training session.
This was a great beginning as it helped me create a metaball shader. However, I wanted to share the location of the spheres at runtime without having to create a hundred nodes (to be exact, 128 for 64 spheres) in material editor, then reference them in C++. The real nightmare was changing the number of spheres. Imaging redoing all of that again.
That’s why I used render targets to feed the spheres’ location to the custom node, which loops over the pixels and sets the spheres’ location and radius based on their RGBA value.
Here’s a method for doing that. I’m sharing it here for anyone who wants to use it, UE’s documentation is not the best. The numbers are all hard-coded, as this was a feasibility test.
void URenderTargetHandlerComponent::StoreSphereLocations(UTextureRenderTarget2D* RenderTarget,
const TArray<FVector>& SphereLocations) const
{
if (!RenderTarget || SphereLocations.Num() != 64)
{
UE_LOG(LogTemp, Warning, TEXT("Invalid render target or sphere location count."));
return;
}
FVector2D DrawSize = FVector2D(RenderTarget->SizeX, RenderTarget->SizeY);
UCanvas* Canvas;
FVector2D Size;
FDrawToRenderTargetContext Context;
UKismetRenderingLibrary::BeginDrawCanvasToRenderTarget(
GetWorld(), RenderTarget, Canvas, Size, Context);
if (Canvas)
{
for (int32 i = 0; i < 64; ++i)
{
int32 X = i % 8;
int32 Y = i / 8;
FVector Location = SphereLocations[i];
FCanvasTileItem TileItem(FVector2D(X, Y),FVector2D(1, 1),
FLinearColor(Location.X, Location.Y, Location.Z, SphereRadius));
TileItem.BlendMode = SE_BLEND_Opaque;
Canvas->DrawItem(TileItem);
}
UKismetRenderingLibrary::EndDrawCanvasToRenderTarget(GetWorld(), Context);
}
}At this point, All the particles were Actors, and I used the UStaticMeshComponent class for particles and their AddImpulse() method to move them around. This approach was good enough to test things out but not efficient, especially considering that I didn’t need most of the UStaticMeshComponent features. I thought of using USphereComponent, but even then, I was using a single CPU thread to process something suited for the GPU.
That’s why I considered the Niagara System (not the Niagara Fluid System). It made sense because I wanted to use GPU while having physics and collisions, and the Niagara System had all of these. I just had to connect user input to the particle system and the particle system to the ray marching shader.
At first, I wanted to calculate the attraction force for each particle, but it was too slow. While I had a lot of “fun” trying to create a Niagara Data Interface, I did not use it and just sent one vector as the mean magnetic center to the system.
Here, I packaged the prototype for iOS and tried it on my phone. It was really slow! And I wasn’t surprised because ray marching is an expensive technique. I tried lowering the sample count to reduce the lag, but it caused weird visual artifacts. To solve that, I added some noise to the sampling distance, which fixes those problems but adds noise to the final result. To embrace (!) the noise, I added a post-processing effect that adds some noise to everything.
I tweaked the parameters more and removed the “asteroids” to get a better visual result, and here I got. This version runs smoothly on an iPhone 15 but it’s still way far from a production-ready state.
While it was a great experiment, I parked the idea here to work on more exciting stuff!
