Breaking01 / AI
PhysICL predicts physical fields without retraining
PhysICL uses simulation examples to predict physical quantities across a new mesh while keeping its model weights fixed.

Researchers announced PhysICL on 13 Nov, a model that predicts physical fields from simulation examples without retraining, according to their paper.
The paper describes a model pretrained solely on procedurally generated tasks. It uses labelled simulation examples to infer physical quantities across an unstructured mesh, a grid with irregular connections, without updating its model weights.
In researcher-published benchmark figures, PhysICL's prediction error on PLAID was lower than the comparison baseline from K = 32 examples. The baseline was the simulation among the supplied examples most similar to the query.
The researchers reported that the same model accepts different numbers of examples and input and output physical quantities. Mesh size and connectivity can also vary within a task.
In researcher-published measurements, peak GPU memory use on AhmedML was 4.27 GiB with K = 64 examples. The corresponding figure for the comparison dataset, 2D_MSHE, was 4.07 GiB.
The paper's abstract does not disclose a method for integrating PhysICL with a game engine.