|
一个简单的流体效果:
https://www.zhihu.com/video/1471428922994425856
工程:
https://github.com/alen-cell/PhysicsEngine
Navier-Stokes equation方程
基于[1]中的论文的算法实现:
描述一个材料点的受力可以由NS方程给出m理解为f = ma 的ma,一个单位质量乘上加速度。
[2]
来自太极 流体力学
关于数学定义:
计算受力
1.pressure
Navier-Stokes equation的第一项便是压强,对于每一个粒子收到的压强不会相同,但是为了使得系统稳定,
将采用一个对称的压强计算
在上一个公式中,计算压强场的梯度之后这个压强的偏差并没有被计算进去。而在计算力之前我们需要计算一个偏移,压强并没有直接被粒子的属性给出,但是可以首先被计算出来:压强的力需要用点所在的位置的密度决定。
ρ0 是静止时水的密度,k是空气含量?根据温度变化而变化
2.Viscosity粘滞力
粘滞力只取决于速度场的偏差而不是绝对的速度,所以计算对称的粘滞力的公式为:
3.表面Tension:
后来再加。。
算法部分:
//ComputeShader kernel定义
private void FindKernels()
{
clearHashGridKernel = computeShaderSPH.FindKernel("ClearHashGrid");
recalculateHashGridKernel = computeShaderSPH.FindKernel("RecalculateHashGrid");
buildNeighbourListKernel = computeShaderSPH.FindKernel("BuildNeighbourList");
computeDensityPressureKernel = computeShaderSPH.FindKernel("ComputeDensityPressure");
computeForcesKernel = computeShaderSPH.FindKernel("ComputeForces");
integrateKernel = computeShaderSPH.FindKernel("Integrate");
DispatchKernel = computeShaderSPH.FindKernel("DrawMesh");
}
核函数
Muller的论文中,给不同的受力不同的核函数,我们使用Wpoly6来做核函数去计算每一个point的物理量(压强和粘滞力除外):
核函数的梯度和Laplacian算子:
梯度
Laplacian
对于收到的Pressure力,我们用Desbrun的Spiky函数描述,因为这个函数在接近中心的时候的梯度不会突然消失,如果用Wploly6的模型,粒子就会在失去斥力在中心点并集结成Cluster。
在计算压强时
时间积分
所以,我们要计算的便是显示计算受力情况,解出收到的粘滞力和重力之和更新中间速度,在求出压强更新速度,最后用速度更新位置。
Debug
要是看ComputeShader的数值,可以将CommandBuffer的值传递回CPU,
用CommandBuffer.GetData()函数:
//Debug.Log(_argsBuffer);
//if (material != null && particleMesh != null)
//{
Vector3[] data = new Vector3[numberOfParticles];
_velocitiesBuffer.GetData(data);
Particle[] dataposition = new Particle[numberOfParticles];
_particlesBuffer.GetData(dataposition);
float[] pressureforce = new float[numberOfParticles];
_pressuresBuffer.GetData(pressureforce);
float[] densities = new float[numberOfParticles];
_densitiesBuffer.GetData(densities);
Vector3[] forces = new Vector3[numberOfParticles];
_velocitiesBuffer.GetData(forces);
material.SetFloat(SizeProperty, particleRenderSize);
material.SetBuffer(ParticlesBufferProperty, _particlesBuffer);
for (int i = 0; i <Mathf.CeilToInt(numberOfParticles/100); i++)
{
//if (pressureforce > 1000|| pressureforce < -1000){
Debug.Log(i+&#34;number Pressure = &#34; + pressureforce);
//}
// Debug.Log(&#34;Velocity = &#34;+ data);
//Debug.Log(&#34;Position = &#34; + dataposition.position);
Debug.Log(&#34;Densities = &#34; + densities);
//Debug.Log(&#34;forces = &#34; + forces);
}
查看先前定义的数值时可以使用FrameDebugger:
稀疏哈希列表
将将网格的大小设置为粒子的SupportRadius的两倍,只需要遍历周围2X2方向的网格。
也就是每个粒子会拥有2x2x2大小的邻接粒子表。
void GetNearbyKeys(int3 originIndex,float3 position, out int nearbyKeys[8])
{
int3 nearbyBucketIndices[8];
for (int i = 0; i < 8; i++) {
nearbyBucketIndices = originIndex;
}
if ((originIndex.x + 0.5f) * CellSize <= position.x) {
nearbyBucketIndices[4].x += 1;
nearbyBucketIndices[5].x += 1;
nearbyBucketIndices[6].x += 1;
nearbyBucketIndices[7].x += 1;
}
else
{
nearbyBucketIndices[4].x -= 1;
nearbyBucketIndices[5].x -= 1;
nearbyBucketIndices[6].x -= 1;
nearbyBucketIndices[7].x -= 1;
}
if ((originIndex.y + 0.5f) * CellSize <= position.y) {
nearbyBucketIndices[2].y += 1;
nearbyBucketIndices[3].y += 1;
nearbyBucketIndices[6].y += 1;
nearbyBucketIndices[7].y += 1;
}
else
{
nearbyBucketIndices[2].y -= 1;
nearbyBucketIndices[3].y -= 1;
nearbyBucketIndices[6].y -= 1;
nearbyBucketIndices[7].y -= 1;
}
if ((originIndex.z + 0.5f) * CellSize <= position.z) {
nearbyBucketIndices[1].z += 1;
nearbyBucketIndices[3].z += 1;
nearbyBucketIndices[5].z += 1;
nearbyBucketIndices[7].z += 1;
}
else
{
nearbyBucketIndices[1].z -= 1;
nearbyBucketIndices[3].z -= 1;
nearbyBucketIndices[5].z -= 1;
nearbyBucketIndices[7].z -= 1;
}
for (int j = 0; j < 8; j++) {
int3 nbcellIndex = nearbyBucketIndices[j];
if (nbcellIndex.x < 0 || nbcellIndex.x >= Dimensions || nbcellIndex.y < 0 || nbcellIndex.y >= Dimensions || nbcellIndex.z < 0 || nbcellIndex.z >= Dimensions) {
nearbyKeys[j] = -1;
}
else {
nearbyKeys[j] = Hash(nearbyBucketIndices[j]);
}
}
}
将每个粒子的临近Particles加入NeighbourList表中:
[numthreads(100, 1, 1)]
void BuildNeighbourList(uint3 id : SV_DispatchThreadID)
{
_neighbourTracker[id.x] = 0;
const int3 cell = GetCell(_particles[id.x].position);
int cells[8];
GetNearbyKeys(cell, _particles[id.x].position, cells);
for (uint j = 0; j < 8; j++)
{
if (cells[j] == -1) continue; // Grid does not contain cell.
const uint numberOfParticlesInCell = min(_hashGridTracker[cells[j]], maximumParticlesPerCell); ;
for (uint index = 0; index < numberOfParticlesInCell; index++)
{
const uint potentialNeighbour = _hashGrid[cells[j] * maximumParticlesPerCell + index];
if (potentialNeighbour == id.x) continue;
const float3 v = _particles[potentialNeighbour].position - _particles[id.x].position;
if (dot(v, v) < radius2) // Use squared length (= dot) instead of length for performance.
{
_neighbourList[id.x * maximumParticlesPerCell * 8 + _neighbourTracker[id.x]++] = potentialNeighbour;
}
}
}
// n. The Neighbouring-list should be n-particles big, each index containing a list of each particles neighbours in radius r.
有了NeighbourList就可以计算压强了,这里的压强我用了Muller论文中的gasConstant来平衡。(大约等于2000f)
因为WCSPH的公式需要小的时间步长。而这里的公式可以达到0.01时间步长不会爆炸。
[numthreads(100,1,1)]//合并
void ComputeDensityPressure(uint3 id :SV_DispatchThreadID)
{
float3 origin = _particles[id.x].position;
float sum = 0.f;
for(int j = 0;j <_neighbourTracker[id.x]; j++)
{
int neighbourIndex = _neighbourList[id.x*maximumParticlesPerCell*8 +j];
const float3 diff = origin - _particles[neighbourIndex].position;
float distanceSquared = dot(diff,diff);
sum += Wpoly6Kernel(distanceSquared);
}
//computeDensity
//TodO presure
_densities[id.x] = sum*mass + 0.000001f;
_densities[id.x] = max(restDensity, _densities[id.x]);
_pressures[id.x] = gasConstant * (_densities[id.x] - restDensity);
}Tips:(若是不维护一个固定的表格,也可以用一个RadixSort实现,但是还不知道排序的时间和这种方式哪个更省,先都探索一下)参考:https://developer.download.nvidia.cn/assets/cuda/files/particles.pdf
进一步解决的问题:
1.粒子在边界收到的压强会小于应该有的压强,导致粒子粘滞在边界。这个需要通过虚粒子的方式解决掉。
2.碰撞。碰撞有两种方式:SDF,固体粒子。而这两种方法在后面都会加入系统之中。因为两种方法其实密不可分,一种重建粒子表面的方法便是建立在SDF之上的。
3.流体的不可压缩性。将使用PBD计算来解决。PositioBasedDynamics。
4.流体和固体的双向耦合。统一用粒子方式解决。
性能:
20000-40000个粒子稳定在160帧左右:
后期加上屏幕渲染方式,固体粒子双向耦合希望可以在60-100 帧上下
资料:
下一篇:基于PBD的流体模拟和刚体耦合
参考
- ^1https://matthias-research.github.io/pages/publications/sca03.pdf
- ^2https://www.bilibili.com/video/BV1mi4y1o7wz?from=search&seid=16843922597556751206&spm_id_from=333.337.0.0
|
本帖子中包含更多资源
您需要 登录 才可以下载或查看,没有账号?立即注册
×
|