Grid-Free Monte Carlo for Time-Dependent Diffusion

A new grid-free Monte Carlo approach now handles time-dependent diffusion instead of only steady-state problems. The method generalizes walk on spheres for pure Dirichlet cases and walk on stars for mixed Dirichlet-Neumann cases to transient heat equations with initial conditions, time-dependent sources, and time-dependent boundary data.

The practical win is that each random walk carries a finite time budget. At every spatial step, the solver samples an exit time; if there is still time left, the walk continues, and if not, it samples inside the domain and evaluates the initial condition. That lets the solver estimate the solution at any requested time directly, without volumetric meshing, sequential time marching, or choosing a timestep.

For developers working on physically based rendering, simulation, or other PDE-driven tools, this is interesting because it preserves the parallel, progressive, output-sensitive behavior of existing grid-free methods while removing temporal discretization bias. The authors also describe kernel sampling and variance-reduction techniques, including a low-bias exit-time sampler and efficient rejection samplers, which are the kind of details that determine whether a theoretically elegant method is actually usable.

The work also includes a way to share walks across multiple target times, which could matter when you need a time series rather than a single snapshot. The exact production use cases remain to be proven, but the direction is clear: less mesh overhead, less manual timestep tuning, and a cleaner path to transient diffusion on...