Tuesday, November 1, 2011

Gas Giants

Moving on from star rendering in my continuing tour of celestial bodies, I thought it would be interesting to have a go at trying to procedurally render a different type of entity – Gas Giants. 

There are four Gas Giants in our own Solar System, Jupiter, Saturn, Uranus and Neptune each very different from the other, but for my first go at rendering one I decided on the largest of these – namely Jupiter – to see how close to it’s distinctive appearance I could manage at real-time rates.

There is a lot of information available about Jupiter (http://en.wikipedia.org/wiki/Jupiter) and Gas Giants in general (http://en.wikipedia.org/wiki/Gas_giant) on sites such as Wikipedia, again the more I found out about the subject the more interesting it became and I strongly recommend a little background reading if it’s a subject that has any more than a passing interest for you.
Procedural Gas Giant generated and rendered with Osiris

Composite Jupiter image from the Cassini-Huygens probe
(NASA/JPL/Arizona University)

Like stars Gas Giants are conveniently smooth for rendering so I decided to use the same ray-tracing approach that I used for stars described in my previous posts.  This is not only efficient in terms of geometry but also produces lovely smooth spherical surfaces at all screen sizes which is important as I want to be able to fly right down near to the surface of my planets.

I also wanted to be able to produce a wide variety of gas giant effects with only a few input parameters so needed a system that could make interesting visual surfaces with more variation than just their colour.

As seen in the images above, gas giants tend to follow a basic structure of having bands of different coloured material at differing latitudes; these are counter-circulating streams of material called zones and belts and trying to simulate such a major feature seemed like a good starting point.  Using the planet space ‘y’ co-ordinate of the pixel being shaded (essentially the latitude) as a lookup into a colour ramp texture is a nice simple starting point for this effect.  In this case I generated the (1 x 2048) sized look up texture by taking a slice out of the above cassini image of Jupiter and filtering it down to the correct size – this gave me a genuine Jovian colour palette straight off-bat, support for other colours is described below.
Basic colour bands from look-up texture
Using Noise

Already there is a basic sense of a Jupiter like planet but of course it’s way too regular and ordered to give a proper idea of the thick turbulent gaseous atmosphere that makes up so much of this type of world.  Adding some degree of noise to the choice of colour from the ramp texture is an obvious choice in this situation, using the same volume noise texture that I used for the star rendering and combining two weighted channels of two octaves gives a far more interesting result
Noise channel #1

Noise channel #2

Noise channel #3

Noise channel #4

Combined noise

Colour ramp rendered with noise

This looks pretty decent from this fairly distant viewpoint but as the camera moves in towards the surface the pattern loses detail as it’s stretched out over more screen space.  Although the naturally soft nature of the surface detail makes this far less unsightly than it would be on a planet surface with more innate contrast, I thought it could be improved by adding additional octaves of progressively higher frequency noise.

The problem with adding higher frequencies however is that you start to get under-sampling artifacts showing up as sparkly noise in the effect when viewed from too large a distance for that frequency - an almost classic example of shader aliasing.  To avoid this two blend weights are calculated based upon a combination of the distance of the planet from the camera and it’s radius.  These two blend weights are passed to the gas giant pixel shader and used to control the contribution of the upper two bands of noise frequencies.  In this way the higher frequency noise blends in smoothly as the camera nears the surface and fades out as it moves further away.  By choosing the blend distances appropriately the transition is essentially invisible.

Close in to gas giant surface without high frequency noise
Close in to gas giant surface with high frequency noise
Silhouette Edges

The interior of the planet is now looking far more interesting but the silhouette edges are letting it down a bit.  Even at 1280x720 resolution the hard edges are unsightly and without any other anti-aliasing currently in use they jar against the smooth colour graduations of the surface.

While the plan is to add anti-aliasing in the future (either MSAA if performance allows or possibly FXAA) I realised there is something that can be done in the shader itself to improve matters and provide some other desirable effects.  Rather than having a hard edge around the sphere it would be far nicer to have a thin band of alpha pixels fading off around the edge to smooth things out.

Fortunately this is quite simple to do, by using a Fresnel term calculated with a dot product of the surface normal and the ray from the camera to the pixel being shaded the alpha value of the pixel can tend to zero as the surface becomes tangential to the view direction.  The only trick required is to control the amount of Fresnel to use to give a consistent on-screen width of just a couple of pixels for the fall off zone – without this the alpha zone would grow and shrink in size as the camera changed distance.  The Fresnel zone is calculated once in C++ and passed to the shader as a constant.

Planet silhouette edges without Fresnel based alpha - aliasing very evident
Planet silhouette edges with Fresnel based alpha - much smoother
A handy side effect is that by clamping the distance factor a larger fade out zone can be used when the camera nears the planet surface producing a completely non-scientific but rather nice atmosphere effect to soften the horizon at these close-up ranges - this can be seen in the high frequency noise planet close up images above.


Adding noise certainly made the surface of my gas giant more interesting but I wasn’t satisfied, it still wasn’t quite what I had hoped for.  Reading about gas giants and looking at the Cassini image at the top of this post it’s quickly becomes obvious that a primary feature of these planets are large numbers of cyclones and anti-cyclones ranging in size from barely perceptible to ones such as the “Great Red Spot” on Jupiter several times larger than the Earth itself combining to make for a highly turbulent surface.

Modelling planet scale cyclones accurately is a computationally expensive proposition beyond the capabilities of a real time system but there is something that can be done on a purely visual level to give at least a nod towards their presence.  To do this I decided to model a system where a potentially large number of cyclones could be represented as a set of cones emanating from the centre of the planet each with their own radius and rotational strength.  In the pixel shader the point being shaded is tested against this set of cones and should it fall within one is then rotated around that cone’s axis by the cone’s rotational strength scaled by the points distance from the cone’s central axis.

Testing a point against potentially hundreds of cones in the pixel shader is however a whole heap of processing so an efficient way to do so is required.  The method I decided upon was to store the cone information in a relatively low resolution cube map texture that could be sampled using the normal of the planet sphere at the point being shaded to determine which cone the point in question fell within, if any.  This way a single texture read provides all the information without having to iterate over each cone individually in the shader.

The cyclone cube map texture uses an uncompressed 32bit format encoded as follows:
  • Red : X component of the unit length axis of the cone that encloses or is closest to this texel
  • Green : Y component of the unit length axis of the cone that encloses or is closest to this texel
  • Blue : Normalised rotational strength of the cyclone whos cone this texel is affected by
  • Alpha : Normalised radius of the cyclone whos cone this texel is affected by, the sign of this value represents the sign of the Z component of the unit length cone axis that is computed in the shader
The format is set up so each byte is interpreted by the GPU as a signed normalised number so the values -128 to +127 are stored in the texture bytes which then arrive in the range [-1, +1] when the texture is sampled in the pixel shader.

The normalised rotational strength of the cyclone and normalised radius values encoded in the texture are scaled in the shader based upon minimum and maximum values passed to it in shader constants.

A little experimentation showed that between 100 and 200 cones of sensible radii could be effectively encoded using a cube map only 128x128 on each side, no two cones being allowed to be within one texel’s worth of angle to avoid filtering artifacts.  A benefit of storing axis and radius information in the cube map rather than actual per-texel offsets is that oversampling artifacts in the texture sampling are avoided as the distance calculations in the shader are operating on the same axial values for every pixel affected by that cone – the down side is that each pixel can only be affected by a single cyclone but I am happy to live with that for now.

Colour coding the pixels based upon the cone axis they are closest to and therefore affected by produces what is essentially a spherical Voronoi diagram of the points where the central axes of each cone penetrates the surface of the sphere
Voronoi diagram showing area of influence of each cyclone

Colour coding the pixels instead with a greyscale value indicating the distance from the central axis of their nearest cone produces the following effect
Position and strength of each cyclone
Clearly indicating the size and position of each cyclone.  Finally, using a combination of this distance value and a global maximum rotational strength supplied to the shader in a constant to rotate each sample point around it’s closest cone axis produces the final cyclone effect:
Final Gas Giant with cyclones
A number of cyclones of different size and strength are easily visible here – it’s not perfect but I think it’s a worthwhile addition that makes the overall effect more interesting.

Colours and Bands

The final step is to allow the shader to produce more varied effects for the potentially large number of gas giants I want my procedural universe to be able to contain.  There are two additional parameters that I’ve added to help support this increase in variety.

The first is a simple scale value that is used to control the mapping of latitude values onto the colour ramp texture.  By increasing the scale the width of the bands can be reduced and their number increased while decreasing the scale of course produces the opposite effect – fewer, wider  bands
Gas Giant with twice as many bands
Gas Giant with half as many bands

The second is a colour remapping post-process for taking the RGB generated by the shader and altering it to add variety.  The manner I thought simplest to do this was to pass three vectors to the shader that are multiplied by the red, green and blue shader colour values immediately before the shader returns.  For example passing (1, 0, 0), (0, 1, 0) and (0, 0, 1) would leave the colour unchanged as the red output is just the red input, the green output just the green input and so on.

Passing different values however completely changes the result allowing the final red, green and blue components that end up on the screen to be an arbitrary mix of those from the shader proper.  Although any range could be used some care needs to be taken to keep the general luminosity in a sensible range or we end up with completely black or pure white planets.  A degree of over-brightening can however produce some funky alien worlds.
Colour remapped gas giant
More dramatically colour remapped gas giant
Further variation could be achieved by using a different noise  texture, different noise octave scales and weights or a different colour ramp texture but for now I think this is sufficient.

Future work

Although the cyclone system was developed for gas giants, there is no reason it couldn't also be applied to other bodies with atmospheres.  The clouds on my Earth style planets for example are currently read from a fBm based cube map but it would be interesting to see how they look with cyclones applied – something else to put on my already lengthy TODO list I think.

This post is already in great danger of suffering from TL:DR so I’ll wrap it up there, my parting shot is this image from the surface of one of the Earth type planets in my solar system that shows the gas giant rising over the horizon.  Always more to do, but I’m quite pleased with the result so far.

Gas Giant viewed from the surface of a nearby Earth-like planet

No comments:

Post a Comment

Comments, questions or feedback? Here's your chance...