I recently had the pleasure to have an insightful discussion with Ms. Fatma Kocer-Poyraz of Altair about trends in HPC, data analytics and of course, AI. I thought I'd share some of the highlights of her views in a brief blog. You can listen to the full discussion here.
Simulation software is revolutionizing the world of product design. By simulating real-world environments, engineers can quickly test and iterate designs without much of the time and resources traditionally required to build actual physical prototypes and testing environments.
Of course, it is also not a cure-all. Developing and implementing simulations can themselves be costly and out of reach for some product design teams.
"Building simulation models sometimes takes so much work," says Fatma Kocer-Poyraz, VP of Engineering Data Science at Altair.
In particular, for businesses building large-scale machinery, it can be difficult to deliver feedback on time to product designers when working with simulations.
"I have customers that say they can only afford to do one simulation, and imagine having to make a decision -- yet an important decision -- with just one data point," Kocer-Poyraz says.
Given these challenges, Altair customers are asking for ways to speed things up. Altair builds software for simulation, HPC, data analytics and AI. Headquartered in Troy, Michigan, the company serves the automotive industry, as well as the aerospace, defense and electronics industries, among others. It's also recently expanded into finance and insurance.
There are steps organizations can take to get the most out of their experiments, Kocer-Poyraz says, such as parameterizing designs, running synthetic data generation, training machine learning models and optimization. Sometimes, however, these steps only go so far.
Altair is trying to address this problem with Altair Physics AI, a new tool that uses historical simulation data to make decisions for new projects.
The tool uses geometric deep learning to "tap into... existing simulations to train machine learning models," Kocer-Poyraz explains, "which means that we are bringing second life to the simulation that [a company] already paid for, that they... used at one time to make a decision in the past. [We] reuse it to make fast decisions for your future programs. And I think that's invaluable because that means you will use all your past simulations to be able to find high potential designs. So we're not saying that we're replacing the simulation process, you'll always be relying on it. But we're helping the user to pick which designs to spend stimulation resources on."
While Altair is a software company, "I have to pay attention to hardware because we are attempting to speed up the solution processes, and hardware is a piece of that puzzle," Kocer-Poyraz says. Altair has a history of using many AMD offerings, including AMD EPYC-processor based servers.
While Altair was founded nearly 40 years ago, the company has built up a team of engineers and data scientists dedicated to AI and machine learning applications in engineering. The team comprises around 20 engineers with backgrounds in specialties such as aerospace and mechanical engineering.
The team established three guiding principles, Kocer-Poyraz says, that help it assess the value of AI/ML for engineering applications: "We said we will be looking at replacing repetitive, non-value-added tasks with machine learning," she explains. "We will be looking at emulating expert decision making, and then we'll be looking at augmenting the engineers' existing capabilities."
The synergy of hardware and software evolving in unison is compelling. The guiding principles Fatma shared on the podcast are well worth consideration for many companies working to determine how and where to apply AI techniques to boost their efficiency, quality, innovation and productivity.
