In a Data Science Connect webinar with HPE and NVIDIA, HPE Federal AI Chief Technologist Steve Heibein explained how probabilistic AI and physics-based simulation work together. He covered COVID research, weather, NASA's AI-designed parts and why rising GPU power is pushing data centers to liquid cooling.
Key takeaways
- Two complementary methods. AI is probabilistic while HPC simulation is physics-based and mathematical, Heibein said.
- AI narrows, simulation confirms. In COVID research, AI cut the candidate spike-protein shapes to a subset that physics modeling could solve.
- Generative design is real. NASA used a large language model to design 3D-printed parts from load and material requirements.
- Same infrastructure for both. AI and simulation each need large, dense accelerated clusters with fast interconnects and storage.
- Air cooling is running out. With AI servers drawing 12 to 15 kilowatts, many racks can hold only a few, pushing data centers to liquid cooling.
High-performance computing and AI are converging, and the combination lets organizations answer hard questions faster and in greater detail, Steve Heibein, Federal AI Chief Technologist at Hewlett Packard Enterprise, said in a Data Science Connect expert webinar brought to the audience by HPE and NVIDIA. Brian Mink, co-founder and president of Data Science Connect, introduced the session.
Heibein explained how probabilistic AI and physics-based simulation complement each other, where the pairing is already in use, and what it means for compute, power and cooling.
What is the difference between AI and traditional HPC?
Heibein used a Russian dolls analogy for the terms. Artificial intelligence is the broad category, which once meant rules-based expert systems, such as a decision tree modeled on how a doctor triages a patient. Machine learning sits inside it, with algorithms that learn from data and improve as they get more of it, such as a model trained on large sets of chest X-rays to spot lung cancer. Deep learning, which uses neural networks, is a specific method of machine learning. Generative AI finds patterns in large bodies of text, genomes, chemicals or video and uses them to generate new content.
Traditional HPC is different in kind, he said. It runs physics-based modeling and simulation, such as finite element analysis and computational fluid dynamics, built on equations like Euler's and Bernoulli's. "As opposed to AI, which is probabilistic," he said, HPC "is mathematical in nature."
How does AI augment HPC simulation?
The two approaches work best together, Heibein said, with AI narrowing the search space so expensive simulations run only where they matter.
His main example was COVID-19 research on U.S. Department of Energy supercomputers built by HPE. AI narrowed the potential shapes of the virus's spike protein from what would have been an effectively unlimited number to a smaller subset. Physics-based modeling then determined the exact shape, which was handed to pharmaceutical companies looking for a drug that would bind to the receptor.
He described similar pairings in other fields:
- Weather: AI makes physics forecasts hyper-local, down to a neighborhood, and can focus detail along a storm's predicted path. He cited NVIDIA Earth-2, a digital twin of the planet that combines AI, visualization and simulation to predict events hours ahead and run what-if scenarios.
- Oil and gas: AI predicts promising fields so exploration starts in a better place and needs less simulation.
- Manufacturing: AI runs quickly through shape simulations before physics validation, and supports predictive maintenance, defect analysis, supply chain and pricing.
- Energy: Power distribution is largely a routing problem where AI can aid the physics.
How is generative AI being used in engineering design?
Heibein described a NASA project that used a large language model to generate parts from requirements such as material, load, torque and mounting-hole positions. The parts are 3D printed in aluminum, stainless steel or titanium, preparing for missions where astronauts can't wait for a replacement part to be shipped.
The resulting shapes are ones that "as a human we probably wouldn't have come up with," Heibein said. They meet the load and temperature requirements with the least material.
He also pointed to code copilots refactoring government COBOL systems, since most COBOL programmers have retired.
What infrastructure do AI and HPC workloads need?
Simulation users want higher resolution, higher fidelity and larger data sets, while AI teams want larger models and shorter training windows, Heibein said. Both need the same thing: large clusters of dense, purpose-built accelerated compute with high-bandwidth, low-latency interconnects and fast, high-throughput storage. "Really this has become a supercomputing problem," he said.
He outlined three ways HPE and NVIDIA deliver it: packaged AI applications that are configured, not coded; a turnkey private cloud AI appliance with a three-step retrieval-augmented generation setup (pick a model, point it at your data, choose an interface); and custom AI factories ranging from 16 GPUs to 16,000.
Why are AI data centers moving to liquid cooling?
Power is the constraint. "A thousand watt GPU needs a thousand watts of cooling," Heibein said. Older data centers may allocate as little as 10 kilowatts per rack and modern ones 25 to 50 kilowatts, while a single AI server can draw 12 to 15 kilowatts. That leaves room for only a few servers per rack unless the facility moves to liquid cooling. Meanwhile, multi-layer chips need lower case temperatures so the lower layers stay cool.
He listed the options: retrofit an air-cooled data center, which is often impractical; build or buy a new one; colocate with providers that already offer liquid cooling; or move workloads to a cloud such as HPE's, which he said is liquid cooled and runs on renewable energy. HPE also builds modular data centers. Heibein said one built in England cut a typical three-year data center build to under nine months.
More AI infrastructure sessions in Data Science Connect webinars
Data Science Connect's 2027 calendar includes 24 free live webinars in four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Technical sessions like this one cover the infrastructure behind enterprise AI. See the webinar schedule.
Frequently asked questions
What is the difference between AI and HPC?
Heibein said AI is probabilistic, predicting likely outcomes from data, while traditional HPC runs physics-based simulations built on known equations.
How can AI speed up HPC simulations?
AI can narrow the range of possibilities, such as candidate protein shapes, so compute-heavy physics simulations only run on the most promising subset.
Why do AI data centers need liquid cooling?
Heibein said a 1,000-watt GPU needs 1,000 watts of cooling, and AI servers can draw 12 to 15 kilowatts in racks often allocated 10 to 50 kilowatts, which air cooling can't sustain at high density.
What are the options for data centers that can't support liquid cooling?
Heibein listed retrofitting, building or buying a new data center, colocating with liquid-cooled providers, moving to a liquid-cooled cloud, or deploying a modular data center.
Watch: Augmenting HPC Workloads with AI: Enhancing Simulation, Efficiency, and Workflow Automation on YouTube
Event: Webinars · More coverage: Webinars insights
