Despite these benefits, the ML training process can be a process of interest. According to Ilya Tolchinsky, Principal Product Manager at ANSYS AI, Simai users must start with over 30 simulations to train a reliable model. Additionally, training can take between 48 hours and 5 days depending on the complexity of the model.
This is a considerable amount of time commitment, but the finished ML model becomes a permanent asset engineer that can be used to explore multiple designs “What-ifs” in minutes, rather than the time/day it takes to run multiple traditional CAE simulations. According to ANSYS, Simai users can expect a well-trained ML model to return results 20-200 times faster than traditional solvers.
The speed is great, but it doesn't make much sense if there's no accurate results. Note that just as solver-based simulations are approximations to real-world physics, ML training models are approximations to solver results. In short, it's an approximation.
Read more: Expert Pointers and Insider Strategy for Peak CAD Performance
That said, Simai has two built-in methods to help users achieve acceptable accuracy with ML models. Like other ML approaches, Simai stores a percentage of simulation data to test the trained ML models for accuracy.
Additionally, we provide a reliability score for each scenario investigated. The closer a particular ML prediction is, the more similar it to the known solutions contained in simulation training data, the higher the score. The company says that Simai models trained with a sufficient amount of diverse training data can return predictions that match simulation results within a few percentage points.
A good start to future simulations
The application of machine learning to simulation has accelerated rapidly over the past five years in terms of power and ease of use, but it is still too early for the integration of these technologies. Given the current trajectory of the industry, it's not difficult to imagine a future that isn't far enough to be linked with CAD design, traditional simulations, and ML, all offered as comprehensive software (SAAS) options.
Leveraging cloud-based storage and enterprise-grade GPUs, users can design, prepare, and run multiple simulations of that design online, simultaneously, and package those simulations into datasets used to train one or more ML models. It's definitely easier than that. Still, it would be interesting to see how things develop over the next five years.
Editor's Note: This article is a companion work to Mike McLeod's feature “Crunching the Numerics: Computer Aided Engineering.” To sign in, read his complete analysis of how simulation software readers can over-accelerate FEA/CFD using machine learning.
