Collapse of a granular column, aspect ratio near 4, released from rest and
recorded live from the browser client. The two runs differ in one configuration value, the
Coulomb friction coefficient. Left, friction: 0.0: the material carries no yield
stress and spreads until it reaches the side walls. Right, friction: 0.5,
representative of dry sand: the front arrests at a median radius of 0.55 m against 0.87 m, and
the deposit is 0.043 m deep at its median against 0.026 m. Real time, 60 frames per second.
The frictionless deposit reaches the side walls, so
the domain and not the material sets where its front stops.
1.95×real time, 32,768 grains, one RTX 2080
1.38sper-grain predictability horizon
2191×growth of a 50 µm perturbation
0.153rollout error at 2 s, vs 1.132 persistence
An AI world model trained on a dynamics simulation of granular material,
together with the simulation that produced it. The simulation advances 32,768
interacting grains faster than real time on one GPU, streams its complete state
to a WebGL2 client, and visualizes it instantly with no post-processing step.
The world model observes a coarse field of that state, predicts the next one,
rolls itself forward, and decodes the result back into particles the solver will
accept. This project integrates both halves: the dynamics simulation that a
learned model is usually asked to stand in for, and the learned model itself,
scored against the real answer on held-out trajectories. Because chaos puts a
floor under any prediction of this material, that floor is measured here too,
and the model is reported against it.
Raised the accuracy of AERO-F, a 3D distributed-memory C++/MPI fluid dynamics
solver, by fitting the finite volume dual mesh onto fluid interfaces, resulting in
convergence order 2.00 against 0.99 for the baseline method, reducing mass
conservation error by 10x
Enforced a discrete geometric conservation law as a constrained least-squares
projection, eliminating unknowns to reduce an ill-conditioned system to a symmetric
graph Laplacian with 6x fewer variables
Augmented the level set reinitialization with a mass conservation constraint, and
showed the resulting matrix has null space dimension equal to the number of connected
graph components, reducing the system to a scalar
Collaborated on a student development team to create Quail, a lightweight Python
implementation of the discontinuous Galerkin method to enable teaching and prototyping,
obtaining a 30x reduction in compute cost using NumPy tensor operations, and coauthored
a journal paper in SoftwareX
NASA Ames Research Center — Aerothermodynamics Intern
Implemented a simplified physics model into NASA fluid simulation software to reduce
the dimensionality of a 3D model to 1D, reaching a 1000x reduction in total runtime
Implemented grid coarsening and interpolation techniques into NASA aerothermodynamics
software, leading to a 4x reduction in compute cost of coupled flow and radiation
calculations
Publications
Lasemi, A., & Farhat, C. (in preparation). Higher-order multifluid interface
representation with dual-mesh motion and DGCL-preserving numerical fluxes.
Journal of Computational Physics.