Crucible2.5D

Crucible2.5D is a material laboratory built around three simulation methods: MLS/APIC MPM, weakly compressible SPH, and Eulerian fluid flow. Its mechanics remain planar; the 2.5D name reflects selected thickness and support approximations. The point is to combine their strengths in scenes that can be edited and inspected while they run: liquid jets, tearing solids, heat-driven processing, ferrofluid, fire and blast.

I built Crucible2.5D as a 2D testbed for advanced simulation techniques I encountered in SIGGRAPH papers and demonstrations, and wanted to understand by implementing and combining them. Keeping the world planar makes local interventions, fields and failures easier to see, and leaves more of an interactive frame budget for material state and solver coupling. Particle count, grid resolution, field iterations and stiffness-dependent substeps still determine what fits that budget.

The recordings below show desktop sessions: 35 complete takes, just over eighteen minutes in all. Start with living material growing over a support, then compare food surfaces and follow heat, tearing, magnetism and blast in their own sections. Alternate takes are grouped under expandable headings; playback stays at its original speed with audio available.

The laboratory overview, finished food surfaces and representation comparison introduce the whole project. MPM, SPH and Eulerian flow then have their own sections. The coupling section traces what the solvers exchange, where each quantity is stored, and when it is applied. Later material extensions build on that foundation.

The native application combines C++ orchestration, OpenGL compute passes and an interactive material editor. A scene can contain deformable or flowing particles, surrounding air, analytic terrain, temperature, damage and external fields. Not every combination is a validated multiphysics model. Some are deliberate visual experiments, and others have specific numerical support that can be traced through the pipeline viewer and source.

The project grew through hands-on implementation: write a material law, build a revealing scene, inspect a failure, then retain an alternative when it answers a different question. Some changes repair a specific exchange or scale error; others trade physical scope for a more expressive interactive model. Keeping those branches visible makes the engineering history part of the experiment.

Food skin is the refined presentation of the same simulated material. The original density and particle views expose structure for inspection; the finished surface adds smooth casing, golden bun crust, porous crumb, cheese, oil and a distinct interior at torn edges. The comparison below shows what this finishing layer contributes. Every capture comes from the native application.

Material coordinates keep the small texture attached as an object stretches. Comparing the original shell with newly exposed boundaries gives a torn edge a different interior appearance. Temperature and stored cooking history drive the live look; separate cooking, char and pore-detail controls let a recording explore appearance without changing the solver.

Compare the diagnostic surface with the finished appearance by dragging the divider. Both captures use the same paused geometry, temperature and material history. Rounded lighting supplies the impression of depth; the simulation supplies the silhouette and material state.

A handful of beads and a drop of water can both be drawn as particles, but they answer different mechanical questions. Independent particles carry position and velocity; contacts keep them from overlapping. Water also needs a response to local compression. Jelly needs that response plus a memory of how the material was stretched or sheared.

In SPH, a particle is a sample of a continuous material. Neighbor kernels estimate density, compression creates pressure, and viscosity responds to relative motion. The native pressure law is weakly compressible: it penalizes excess density rather than enforcing exact incompressibility. In MPM, material points retain deformation and history while a temporary grid mediates forces. That stored deformation lets an elastic body push back toward its rest shape.

So particles are not the limitation: SPH and MPM both use them. Molecular dynamics, spring networks and particle constraints can also produce fluid or solid behavior with suitable interactions. The missing ingredients in a simple ballistic/contact demo are the constitutive law, collective spatial response and—in a solid—persistent strain. MPM can represent a fluid too; its material branch deliberately removes elastic shear memory.

Crucible2.5D asks how material moves through large deformation, contact, flow and changing topology. Modal2D asks which small vibrations a particular object supports, and what a strike will sound like. Those goals favor different numerical work, even though both start from mass, internal forces and boundary conditions.

Modal2D assembles a mass matrix and stiffness matrix from a mesh whose geometry, thickness, material and supports matter to the instrument. Its eigenmodes turn that spatial problem into a compact set of oscillators. A small frequency error can mean an audibly mistuned note, so the shape model, discretization and convergence matter—not simply whether the motion looks convincing.

FEM is possible for large-deformation physics too; it needs suitable nonlinear, contact and topology handling. MPM itself uses a grid basis and a weak-form discretization with moving material samples. This is a choice of representation and workload, not a rule that one project needs accurate physics and the other does not.

The SPH solver uses weakly compressible pressure, viscosity, local surface-tension terms and additional velocity smoothing. Neighbor structure determines which particles exchange these quantities. Signed-distance collision handling and continuous collision checks help keep fast particles from simply passing through boundaries.

A hashed cell structure finds nearby particles. Density is accumulated with a normalized 2D cubic spline of support 2h; the force stage combines pressure gradients, viscosity and interface terms. XSPH then smooths relative particle velocities. The Codim Threads branch detects a local tangent and thinness, adapting the density treatment when a fluid bridge becomes line-like.

The sand-castle and jet-style scenes are useful because their expected interaction is legible: the liquid should transfer momentum to the deformable material rather than merely overlap it. The current SPH-to-MPM path records a reaction impulse, scatters it to a persistent grid and lets the MPM side consume it in the following update. This is a concrete return path, improving on a one-way visual interaction.

The air route uses a staggered velocity grid with pressure projection, advection and buoyancy-related forcing. Smoke, temperature, fuel and oxygen fields add visible and reactive quantities to that motion. The pressure solve uses an iterative grid method, so its iteration budget and boundaries are part of the experiment.

A MAC grid separates cell-centered scalars from velocities on faces. Advection transports the fields; force passes add buoyancy and selected injections; a pressure-potential solve repairs the velocity divergence; boundary passes enforce no-through-flow at solid faces. Temperature diffusion and the oxygen/fuel stages add further field updates. This is a full solver family, with smoke serving as a visible tracer of its motion.

Combustion-inspired behavior includes oxygen limitation, fuel accumulation, smoldering and a backdraft scenario. The chamber demonstration makes the causal sequence readable: a restricted environment can retain fuel while oxygen is scarce; changing access to oxygen changes the reaction. These are authored reduced rules rather than detailed chemical kinetics or a validated fire-safety model.

An open flame can shed hot gas into its surroundings. The same source inside a sealed pocket has to interact with the enclosure: energy and gas accumulate, the walls carry a load, and a breach changes the possible escape paths. The interesting behavior is the transition from storing an effect to releasing it through a newly available opening.

The native sandbox builds coarse occupancy and outside-reachability fields. A cell is enclosed when the outside cannot reach it through the current mask. Trapped pressure, heat and fuel-related state evolve alongside the material; cracks and moving matter can change that connectivity. The mask is deliberately coarse and thresholded, so a visually tiny gap may not yet be an open cell path.

Airtightness and pressure are different questions. Connectivity identifies a possible escape route; conductance, stored state and the release model determine the resulting motion. A pressure projection that tries to remove velocity divergence cannot, by itself, store the thermodynamics of a sealed gas volume. That is why this laboratory has a separate trapped-pressure and venting path.

The default legacy blast combines authored heat, push and smoke responses. A separate selectable Sedov-style mode propagates a geometric front and deposits impulses, heat and air motion as that front crosses the scene. Its strong-front expansion follows a two-dimensional similarity scaling before continuing near a configured acoustic speed. This makes arrival and affected regions inspectable without solving a full compressible flow system.

This distinction is essential. A moving ring and a convincing impulse response do not establish a resolved pressure discontinuity from a compressible Euler solver. The open blast range is best understood as a reduced propagation and coupling experiment, with explicit assumptions about energy, front speed and deposited effects.

Current native shock passes include sampled SDF visibility checks. A front reaching the right radius is therefore only one part of applying an impulse: the route to its receiver also matters. Thin geometry and particle-material shielding remain limitations. The study below isolates arrival between frames and a simple wall test; it does not identify the unshown blast-mode setting in a recording.

The representations do not share one universal pressure or constitutive solve. They exchange specific quantities at specific stages. That explicit structure is the useful technical story: a contact reaction can wait in an accumulator; heat can be mixed on a grid; air can receive smoke and temperature before feeding drag and ambient conditions back to particles.

The SPH jet and thermal-contact benches isolate two of these loops. The diagram below makes the accounting visible before adding spatial sampling and constraints. In the native application, pipeline textures and particle fields expose the same ownership questions inside the larger coupled simulation.

The original mechanical v2 path is optional and is narrower than its “two-way” label suggests. Separating recorded, deposited and actually applied reaction exposes missing terms, receiver-grid losses and velocity clamps. Thermal exchange v2 is a different path and is enabled by default.

The full-contact experiment closes one specific gap: it returns repulsion, restitution and contact damping, deposited at the original contact position. It is available under A/B models → Coupling v2 → Full contact reaction (experimental), with the new option off by default. This improves the local accounting without replacing the existing contact law or erasing receiver-grid losses.

There are two related directional effects. Fiber reinforcement changes the constitutive stress even while a material is intact. The newer AnisoMPM damage branch then changes how positive tensile stress drives weakening: it adds a structural tensor, a matched toughness normalization, retained maximum history, and a nonlocal damage field. Comparing these separately avoids mistaking a stronger fiber for a different fracture criterion.

The anisotropic damage work introduces directional structure rather than relying only on an instantaneous scalar break threshold. A structural tensor carries preferred orientation; a damage-driving history prevents the material from forgetting a previously reached load; a nonlocal term spreads the interaction over a chosen length scale. Residual stiffness keeps the damaged response from becoming a completely unconstrained numerical singularity.

The Aniso Tear Bench compares legacy, isotropic and anisotropic choices in a common scene. This makes fiber direction and damage localization visible as algorithmic differences. The same loading can produce a different fracture pattern because the material model changes how deformation contributes to the driving field.

The nonlocal AnisoMPM fracture modes are selectable and default-off in the audited source. They are currently enrolled in three material branches—Brittle, Tough and Ortho Tear. Adding a material name to a profile table is insufficient without the matching constitutive hooks.

Thermal work includes contact exchange, diffusion, external heating, cooling and material-dependent response. A contact-temperature mixture replaces the older maximum-temperature behavior that could ratchet a system upward merely through repeated contact. The revised transfer is still one part of a larger update containing sources, cooling and bounds.

Material processing extends beyond a single melting threshold. The project explores curing, phase change, gas production, crust formation and different surface/core behavior. Clay, wax, epoxy and baking-inspired presets use these histories to create visibly distinct transitions. Their parameters are authored for the laboratory and are not presented as calibrated constitutive data for industrial manufacturing or food science.

The recordings make three distinctions visible: softening changes the response now; a setting or processing variable preserves a history; and prolonged flow can remove the distortion that would otherwise drive shape recovery. The next section isolates that last distinction. For shell/core materials, setting the exterior first also changes how an expanding interior is restrained.

Porous Growth Comparison isolates another issue: a material should not pull itself back to its unexpanded shape merely because its gas-driven rest area grew. An optional correction now measures elastic stress and tearing against that grown rest shape. Scene 143 compares it with the legacy law and separates both from the scalar thickness controls.

• The thermal bench isolates contact exchange before adding combustion or material conversion.
• The advanced bake bench compares authored surface/core and processing histories.
• Temperature, conversion state and visible appearance should be inspected separately.

Cooling early can preserve enough structure for a deformed body to recover. Heat it for too long and the material can lose the elastic distortion that drove that recovery. Cooling then restores stiffness in a changed shape. I explore that distinction with successive heat and cool brush passes in the recording.

A compatible mechanism exists in the native phase-material branch: heating reduces stiffness and relaxes the deformation gradient. Cooling reverses the phase fraction, but it does not undo that relaxation. The specimen’s exact material is not visible in this take, so the study illustrates this source-backed mechanism without claiming to reconstruct its settings.

Start with the same pool under two interventions. The magnetic probe pulls the material upward into a narrow column while neighboring surface peaks appear; Spring Drag grabs a selected patch and moves it mechanically. The status bar identifies each tool. Switching between them makes field-driven attraction and surface shaping visibly different from a direct local pull, although this hand-operated recording is not a force-matched benchmark.

The surface peaks are the familiar visual cue of ferrofluid: a liquid carrying magnetic particles that responds to an applied field. In three dimensions, the interface can form arrays of spikes across a surface; this planar laboratory shows their profile as peaks and narrow columns. The linked external research video provides that 3D context, including comparisons with real ferrofluid experiments. It is background for the phenomenon, not footage of Crucible or evidence that this simplified model reproduces the same physics.

The ferrofluid branch is built into MPM rather than applied only as a visual force on dots. It resets deformation to a volumetric fluid response, adds viscous stress from the affine velocity matrix, and incorporates coarse directional magnetic stress near the interface. The separate force pass handles field-gradient attraction. Pressure, translation, alignment and surface shaping therefore have distinct controls and failure modes.

The magnetic work combines external fields, material response and iterative field calculations. Temperature and remanence-related state can alter behavior, while simplified soft-iron handling introduces a separate approximation. The ferro-spike bench includes explicit traction and corrugation seeding, so its spikes should not be advertised as an instability emerging solely from a complete self-consistent field solve.

The companion recordings expose field direction while the probe moves, then introduce induced and permanent magnetic responses. Field colors are diagnostic, not temperature or emitted light. A separate fantasy-style path prioritizes visual exploration; the opening sandbox tour also shows living-culture growth, which uses its own procedural field rules.

Twenty-four recording compositions make the existing material laboratory easier to explore. Load the recording category in the native Scene Browser, then choose Load and run or Load paused. Empty magnetic and living-material containers let the act of placing a material introduce the experiment; comparison rows and matched supports keep the remaining scenes legible. P pauses, R resets the take paused, and U hides the interface.

Recording mode also separates the scene heater from its drawn gizmo. Hiding the interface no longer switches off the kiln source in these new scenes. The old linked behavior remains available as an explicit comparison; source temperature and actual activation are different settings.

The new capture controls separate how heat looks from how it evolves. Exposure, saturation, an emission threshold and ironbow/cool–warm/spectral temperature views change the presentation. A thermal reservoir changes the experiment. The optional symmetric-flux route targets one precise diffusion exchange error; it does not make the entire coupled engine energy-conserving.

This inventory groups the system into 25 meaningful capabilities, excluding material-name variants, presets, UI and rendering cosmetics. Twenty-two have substantive implementations; three are stubs or deferred extensions. That is an implementation-scope estimate, not “88% physically correct.” A convincing scene, a complete code path and a calibrated numerical method are different achievements.

The recording set contains 35 native sessions covering refined food surfaces, living growth, field interactions, thermal processing, tearing, explosive scenes and porous depth-support comparisons. Three browser companions add independently checked reduced calculations for recovery, shell/core history and front arrival. These provide clearer behavioral evidence and explanation; they do not replace native convergence or calibration studies.

The three core solvers were exercised together in a new baseline run; stiffness v3 and Contact Heat v2 have documented feature-specific checks. Twelve further capabilities have implemented mechanisms with limited rollout or unresolved calibration, and five are partial coupling or deliberately reduced/artistic systems. These broad classifications describe readiness, not the absence of all testing: the new full-contact variant within the coupling capability now also has a retained, independently reviewed GPU regression.

The baseline check used the existing Debug executable with the current copied shaders and data in an isolated working directory. It reached 1,800 frames in 28 seconds with empty stderr and no high-severity GL or shader errors. Recurring medium-severity transfer warnings remain. This exercised initialization and default-scene progression; it was not a fresh build, a performance benchmark or an all-features acceptance test.

The porous-growth variant adds a separate retained regression: 234 production-shader checks, 24 exact disabled-mode checkpoint comparisons, and four finite 600-frame scene runs. This validates the local growth correction and its compatibility boundary; the broader thermal-processing and 2.5D-support capabilities retain their existing calibration limits.

This project evolved by testing a mechanism, inspecting where it failed, and changing the representation or update order. Some revisions repair a numerical error; others change the artistic objective. A new version does not automatically replace its predecessor, and a selectable implementation is not automatically a calibrated default.

The log groups related revisions by the problem each one exposed. Historical observations remain labeled as such, and source references retain the underlying record. Legacy comparison modes show what a changed stress law, field solve or exchange schedule contributes.