Landmark Simulation Papers

Introduced the semi-Lagrangian advection method for incompressible fluid simulation. Replaces explicit time integration (which requires tiny dt) with a backwards-tracing advection step that is unconditionally stable for any timestep. Combined with pressure projection via Helmholtz decomposition, this is the foundation of most real-time fluid simulators.

Introduced FLIP (Fluid-Implicit-Particle), a hybrid Lagrangian-Eulerian method. Particles carry velocity; a grid handles pressure and divergence-free projection. Each step: transfer particle velocities to grid (P2G), project, transfer back (G2P), advect particles. Produces much less numerical diffusion than pure grid methods.

Made SPH practical for real-time graphics. Introduced the poly6 kernel for density, spiky kernel for pressure gradient (preventing particle clustering), and viscosity kernel for smoothing velocities. Defined the complete SPH pipeline: find neighbours → compute density → compute forces → integrate.

Applies Position Based Dynamics (PBD) to SPH-style fluids. Instead of force-based integration, solves density constraints directly on particle positions. Highly stable, handles large timesteps, compatible with the PBD cloth/rigid-body pipeline. Introduces the tensile instability correction term (artificial pressure).

Solved the stiffness problem of cloth simulation using implicit integration. Previous explicit methods required tiny timesteps to avoid explosion from stiff spring forces. The implicit method solves a linear system of equations each timestep, allowing dt up to 50× larger while remaining stable. Foundation of modern cloth simulation in VFX.

Replaced force-based simulation with direct constraint projection on positions. Unconditionally stable, artist-friendly (constraints are readable geometric conditions), and fast enough for games. Handles cloth, soft bodies, rigid bodies, fluids (PBF), and cables in a unified framework. Used in Nvidia PhysX, Unity, Unreal.

Formalised impulse-based rigid body simulation. When two rigid bodies collide, compute an impulse j = −(1+e)·v_rel·n / (1/m₁ + 1/m₂ + n·(I₁⁻¹(r₁×n)×r₁ + ...)). Apply impulse to both bodies. Handles restitution (e) and friction in a unified framework.

Introduced Sequential Impulse (SI) and warm starting for constraint solvers. The basis of Box2D, Bullet, and most game physics engines. SI iterates over all constraints in sequence, applying corrective impulses, converging in 8–20 iterations. Warm starting reuses previous frame's impulses as an initial guess, reducing iterations needed.

Reduced gravitational N-body simulation from O(n²) to O(n log n). Build an octree (3D) or quadtree (2D) over all particles. For each particle, if a cell is far enough (d/l < θ, typically θ=0.5), treat the entire cell as a single mass at its centre of mass. Otherwise recurse into children.

Introduced spatial hashing as an O(n) alternative to spatial trees for uniform particle distributions. Map a 3D cell coordinate to a hash table index using a large prime hash function. Each cell stores a list of particles. For each particle, check only the 3³=27 neighbouring cells.

Comprehensive survey that established LBM as a mainstream CFD method. LBM operates on a discrete velocity distribution function f_i(x, t) on a grid. Two operations: collision (relax f toward equilibrium via BGK) and streaming (propagate f_i to neighbour in direction i). Macroscopic ρ and u recovered from moments of f.

Showed mathematically that two interacting chemicals (activator + inhibitor) can spontaneously form spatial patterns from a uniform state if they have different diffusion rates. This "Turing instability" explains animal coat patterns (stripes, spots), finger formation, and many biological structures.

Defined the Gray-Scott model — two coupled PDEs with feed (F) and kill (k) parameters that produce a rich pattern zoo depending on the F-k parameter space. The autocatalytic term u·v² creates the activator-inhibitor dynamics. Spots, stripes, worms, coral, mitosis — all from two equations.

Introduced Boids — three simple rules (separation, alignment, cohesion) that produce emergent flocking behaviour. Each boid considers only its local neighbourhood; no global communication. The basis of all crowd simulation, traffic simulation, and swarm AI.

Conway's Game of Life — the canonical cellular automaton. Binary states, local 8-neighbour rules, Turing-complete, self-replicating structures. Demonstrated that complexity arises from extreme simplicity. Influenced generations of simulation, AI (neural CAs), and theoretical computer science.

Extended Boids with steering behaviours: seek, flee, arrive, pursuit, evasion, obstacle avoidance, path following, separation. Formalised as force fields on velocity. Standard vocabulary of game AI movement.