AI for numerical simulation

AI for physics simulation

Vast majority of technologies in the world come to life through months or years of extensive simulation during their development. High-performance computing, parallel processing and GPUs helped push the computation time from months to weeks. With the help of machine learning, we are seeing a reduction from weeks to days. We think that this technology is only scratching the surface. At DimensionLab (opens in a new tab), we are building tools for engineers and researchers to tame the physics of their projects in hours, some even in minutes.

In addition, we're also focusing on makers and creators, who are not trained physicists or mathematicians and want to leverage the extreme effectiveness of AI for physics without dealing with its complexities. Collectively, they make up a cohesive platform we call is a software platform for AI-based numerical simulation. It provides a visual editor for building high-performance AI-based physics simulators (also called surrogate models), tools for training these simulator models in powerful GPU-based cloud servers and easy-to-use AI inferencing and 3D visualization pipelines. There's a paradigm shift happening in the world of physics simulation. Deep learning methods are outperforming even the most optimized classical numerical solvers. The key components of these methods are physics-informed neural networks, variations of Fourier neural operators and deep operator networks. They completely eliminate the need for meshing the geometries.

There is a plethora of ways in which ML/NN models can be applied for physics-based systems. These can depend based on the availability of observational data and the extent of understanding of underlying physics. Based on these aspects, the ML/NN based methodologies can be broadly classified into forward (physics-driven), data-driven and hybrid approaches that involve both the physics and data assimilation.