Physical Simulation

Physical simulation is the branch of computer graphics that computes how things move by applying the laws of physics—or approximations of them—inside a computer.

12 frameworks, from 1986 to today · Computer Graphics · Computer Science

still a working view ended · each leads to its entry below

Overview

Physical simulation is the branch of computer graphics that computes how things move by applying the laws of physics—or approximations of them—inside a computer12. Instead of an animator drawing each frame by hand, a simulator is given a scene: a cloth, a splash of water, a stack of boxes, a character's skeleton. It is given rules about forces, collisions, and constraints, and it steps the scene forward in time, producing motion that obeys those rules. What is at stake is plausibility and control at once. A physically simulated explosion or cape or hair looks right in a way hand animation struggles to match, because it responds to the same causes a real object would. But a simulation that is merely correct is not automatically useful to a filmmaker or a game designer, who need to hit a mark, keep a character upright, or make a splash look dramatic. Much of the field's history is the tension between fidelity and controllability, and between accuracy and speed.

Springs, Particles, and the First Animated Cloth

The earliest graphics simulations borrowed a model from engineering: the mass-spring system. A deformable object is represented as a set of point masses connected by springs34. Each spring pulls its two endpoints toward a rest length with a force proportional to how far it is stretched, and each point's acceleration follows from the forces on it4. The model is simple, cheap, and intuitive, and it became the workhorse for cloth, hair, and soft bodies35.

The problem was numerical. Stiff springs—the kind that keep cloth from stretching like rubber—require very small time steps to stay stable, because an explicit method that computes each new position from the current forces can overshoot and blow up. In the late 1980s and 1990s, graphics researchers built cloth and deformable-object systems on mass-spring networks and wrestled with exactly this stiffness. The practical fix that emerged was to make the integrator implicit: rather than using only the current state to predict the next, an implicit method solves for a future state that is consistent with the forces there6. That requires solving a linear system each step, but it permits far larger time steps. This trade—more algebra per step, fewer steps—became a foundation of the field.

Rigid Bodies, Contact, and the Trouble with Friction

Cloth is soft; the other half of the world is hard. Rigid body dynamics treats objects as undeformable shapes whose motion is fully described by position and orientation7. Simulating a single tumbling box is easy. Simulating a hundred boxes piled in a heap is hard, because the difficulty is not the motion but the contacts: every touching pair imposes a constraint that the objects must not interpenetrate, and friction imposes a further constraint on how they may slide.

The core insight, developed in graphics through the 1990s and 2000s, was to treat contact as a set of inequality constraints solved together, rather than as a sequence of pairwise collisions handled one at a time. Handling contacts sequentially lets errors propagate—push one box out of another and it sinks into a third. Solving them jointly, as a complementarity problem (a mathematical condition stating that either objects separate or a contact force pushes them apart, but not both), gives stable stacks. Friction adds a second layer: a contact can resist tangential motion up to a limit proportional to the normal force, and beyond that it slides8. Getting friction right is what makes a simulated object rest on a slope instead of creeping downhill, and it remains one of the field's most delicate numerical problems.

Constraint Solvers and the Position-Based Turn

The linear systems that implicit integration and contact produce are large and sparse, and solving them exactly is expensive. Graphics took a different route from engineering: iterative solvers that take a few cheap passes and accept an approximate answer. The Gauss-Seidel-style projection used in many solvers updates one constraint at a time, repeatedly, until the violations are small enough to ignore visually. This is fast and robust, and it tolerates the over-constrained, inconsistent systems that real scenes produce.

A decisive shift came with position-based dynamics (PBD), introduced by Matthias Müller and colleagues in the mid-2000s and popularized in a 2007 paper. Instead of computing forces and integrating, PBD moves particles directly to satisfy constraints—a distance constraint pulls two points to their rest separation—and then derives velocities from the position change. It is unconditionally stable, simple to implement, and fast enough for real-time use, which made it the default in games and interactive tools. Its weakness is that stiffness depends on iteration count and time step, so it is not physically accurate; a later refinement, extended position-based dynamics, added a stiffness parameter and better handling of bending and volume. The method's rise marks a philosophical split that persists: force-based simulation aims at physical correctness, while position-based and constraint-projection methods aim at stable, controllable, fast behavior that looks right.

Fluids: Grids, Particles, and the Two Traditions

Liquids and smoke posed a different problem: they have no fixed shape, and their motion is governed by the Navier–Stokes equations, which describe how velocity and pressure evolve in a fluid9. Graphics could not afford the precision of engineering computational fluid dynamics, so it built its own approximations.

The Eulerian approach represents the fluid on a fixed grid: each cell stores velocity and pressure, and the solver advects quantities through the grid and enforces incompressibility by solving for pressure. A landmark was the 1999 "Stable Fluids" method of Jos Stam, which used a semi-Lagrangian advection step—tracing each grid point backward along the velocity field to find where its value came from—and an implicit pressure solve. This made smoke and water simulation stable at large time steps and simple enough to spread widely; it underpins much of the smoke and fire seen in film and games since.

The Lagrangian approach instead carries the fluid as moving particles10. Smoothed particle hydrodynamics (SPH), developed in astrophysics in the 1970s, represents a fluid as particles whose properties are averaged over a smoothing kernel with their neighbors11. Graphics adopted it in the early 2000s for splashes, foam, and free-surface flows, where particles follow the water naturally and grid resolution is not a constraint. SPH's weaknesses—difficulty enforcing incompressibility, sensitivity to particle spacing—drove years of refinement, including the position-based fluid methods that applied PBD's constraint projection to density.

The two traditions competed and then merged. Grid methods handle smooth bulk flow and pressure well; particle methods handle splashes and fragmentation well. Hybrid schemes that couple a grid to particles, and later methods that combine both, became standard for convincing water.

Making Simulation Directable

By the 2000s, simulation was good enough to be trusted with hero effects, and the field's center of gravity shifted toward control. A director does not want a physically correct splash; they want a splash that reads on camera. Art-directable simulation is the set of techniques for steering a simulation toward an intended result without breaking its physical character: adding guiding forces, constraining key particles to follow a target path, or blending a simulated result with a hand-animated one. The same impulse produced physics-based character control, in which a simulated figure is driven to walk, balance, and recover from pushes by controllers that compute the forces needed to track a desired motion while keeping the body upright. Here simulation serves animation rather than replacing it: the physics supplies the weight and balance that make a character believable, while the controller supplies the intent.

The Learned Turn

The most recent era asks whether simulation needs to be solved at all, or whether it can be learned. Neural and learned simulation trains a model on examples of simulated or captured motion and then predicts future states directly, often far faster than a numerical solver. The appeal is speed and the ability to fit complex materials that are hard to model from first principles; the risk is that learned models drift, violate conservation laws, or fail outside their training data.

A parallel development is differentiable simulation, in which the entire simulation is written so that gradients of an output—say, the final position of a cloth—can be computed with respect to its inputs, such as material parameters or control forces12. This lets optimization and machine learning tune a simulation to match a target, closing the loop between the simulator and the animator's intent. Both directions are young and unsettled, and they inherit the field's oldest tension: a simulator that is fast and controllable may not be accurate, and one that is accurate may be neither.

Where the Field Stands

Today's production and research practice draws on all of these layers. Mass-spring and position-based methods handle cloth, hair, and soft bodies in real time; rigid body solvers with joint contact and friction handle destruction and stacking; grid and particle fluid solvers, often coupled, handle water, smoke, and fire; and constraint solvers and implicit integrators sit underneath all of them2. Learned and differentiable methods are being folded in as accelerators and as tools for control and parameter fitting rather than as replacements. The open questions are the old ones in sharper form: how to make simulation both physically faithful and directly controllable, how to keep stiff and contact-rich systems stable at interactive rates, and how to trust a learned model enough to put it in a pipeline where a wrong frame is a visible failure.

References

  1. Simulation (Classification and terminology), Wikipedia
  2. Physics engine, Wikipedia
  3. Mass-spring-damper model, Wikipedia
  4. Spring system, Wikipedia
  5. Soft-body dynamics, Wikipedia
  6. Backward Euler method, Wikipedia
  7. Rigid body dynamics, Wikipedia
  8. Friction (History), Wikipedia
  9. Navier–Stokes equations, Wikipedia
  10. Lagrangian and Eulerian specification of the flow field, Wikipedia
  11. Smoothed-particle hydrodynamics, Wikipedia
  12. Differentiable programming, Wikipedia

The frameworks of Physical Simulation

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1986-2005

    Mass-spring-damper model

    How do springs and point masses make cloth move?

    Fly to its tower in Nousfarer

  2. 1990-Present
  3. 1995-Present

    Constraint Solvers and Projection Methods

    How can many constraints be satisfied cheaply and stably?

    Fly to its tower in Nousfarer

  4. 1995-Present

    Implicit Integration

    How can stiff motion be stepped forward without blowing up?

    Fly to its tower in Nousfarer

  5. 1999-Present

    Eulerian Grid Fluid Simulation

    How does smoke or water flow on a fixed grid?

    Fly to its tower in Nousfarer

  6. 2003-Present

    Lagrangian Particle Fluid Simulation

    How do moving particles reproduce splashes and foam?

    Fly to its tower in Nousfarer

  7. 2005-Present

    Art-Directable Simulation

    How can a simulation be steered to a wanted result?

    Fly to its tower in Nousfarer

  8. 2005-Present

    Physics-Based Character Control

    How can a simulated figure walk and keep balance?

    Fly to its tower in Nousfarer

  9. 2007-Present

    Hybrid Grid-Particle Fluid Coupling

    How can grids and particles be combined for convincing water?

    Fly to its tower in Nousfarer

  10. 2007-Present

    Position-Based Dynamics

    How can positions be moved directly to satisfy constraints?

    Fly to its tower in Nousfarer

  11. 2018-Present

    Differentiable Simulation

    How can gradients flow through a whole simulation?

    Fly to its tower in Nousfarer

  12. 2018-Present

    Neural and Learned Simulation

    Can motion be predicted by learning instead of solving?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Mass-spring-damper modelRigid Body Dynamics with Contact and …Constraint Solvers and Projection Met…Implicit IntegrationEulerian Grid Fluid SimulationLagrangian Particle Fluid SimulationArt-Directable SimulationPhysics-Based Character ControlHybrid Grid-Particle Fluid CouplingPosition-Based DynamicsDifferentiable SimulationNeural and Learned Simulation

Beyond the field

Where Physical Simulation's ideas came from in other fields, and what they led to there: the strongest 12 of 39 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Physical Simulation is a planet in Nousfarer

Its city has a tower for each of these 12 frameworks. Land, walk in, and climb one concept at a time.

Land on Physical Simulation

Related fields

Fields its ideas are linked to

More in Computer Graphics

See where Physical Simulation sits on the Map of Human Knowledge