Artificial intelligence can now model increasingly complex physical systems with remarkable accuracy. But Syed Bahauddin Alam believes one of the most important challenges begins after an AI model works: Can it actually be deployed where the physical system operates?
Advanced energy systems create almost the opposite conditions assumed by today's largest AI models. Sensors may be sparse, degraded or impossible to install. Operating conditions and geometries change. Computing resources close to a facility are limited. And in safety-critical systems such as advanced nuclear reactors, AI must meet stringent requirements for latency, reliability, uncertainty and energy consumption.
Alam, an assistant professor of Nuclear, Plasma, and Radiological Engineering at the University of Illinois Urbana-Champaign and an affiliate of the National Center for Supercomputing Applications, has received a National Science Foundation Faculty Early Career Development, or CAREER, Award for his project, “Energy-Efficient and Compact Foundation Model for Universal Virtual Sensing in Dynamic Energy Systems.”
The project brings together foundation models, virtual sensing, brain-inspired AI, neuromorphic computing and hardware-aware AI design around a single challenge: deployability.
“Accuracy is necessary, but it is no longer enough,” Alam said. “For energy systems, one of the biggest remaining AI challenges is whether intelligence can actually operate where the physics is happening, under real constraints on sensing, power, latency, hardware and reliability.”
Moving beyond component-specific AI
A major limitation in engineering AI is fragmentation. One model may be trained for a heat exchanger, another for a coolant channel, another for a structural component, and another when the geometry or sensor configuration changes.
Alam's CAREER project proposes a different paradigm: a cross-component foundation model that learns transferable physical representations across multiple systems and functions as a common virtual observer rather than a collection of independently maintained AI models.
“A real energy system is not one problem,” Alam said. “If every component needs its own large model and training pipeline, AI can become operationally unaffordable long before it becomes technically wrong.”
The ambition is shared physical intelligence across an engineered system.
Energy and AI hardware cannot be afterthoughts
Another major deployment barrier is frequently overlooked: the energy cost of AI itself.
Large foundation models generally assume abundant computation. Physical infrastructure does not. Alam's project therefore treats energy consumption, memory movement, latency and hardware constraints as design variables from the beginning, rather than optimization problems addressed after the model has already been developed.
This is also where the project connects directly to AI hardware and neuromorphic computing.
Alam's group is developing neuroscience-inspired neural operators and variable-spiking-neuron architectures based on sparse, asynchronous and event-driven computation. Preliminary work reported in the CAREER proposal showed a prototype achieving comparable accuracy to a conventional Fourier neural operator while reducing measured inference power on an edge platform from 4.2 watts to 0.6 watts.
The group is also pursuing hardware-aware AI architectures and algorithm-hardware co-design for low-power, event-driven and neuromorphic platforms.
“The brain shows that powerful intelligence does not require dense computation everywhere, all the time,” Alam said. “But sparsity on paper is not efficiency on hardware. The challenge is making algorithmic efficiency translate into measurable deployment efficiency.”
National experiences shaped the research agenda
The CAREER project represents the convergence of several years of Alam's work across AI, advanced computing and nuclear energy.
He served on the 11-member National Academies committee on Foundation Models for Scientific Discovery and Innovation, was identified as a national AI leader and expert in Illinois' response to the White House AI Action Plan, and was selected for the National Academy of Engineering's U.S. Frontiers of Engineering, representing Compute Challenges for Artificial Intelligence.
His work has also been recognized through the DOE Distinguished Early Career Award, ANS Nuclear News 40 Under 40, and HPCwire Editors' Choice Award in Energy.
“These experiences kept pushing me toward the same question,” Alam said. “What should trustworthy, energy-aware AI look like when it has to operate in the physical world? The National Academies work especially pushed me beyond benchmark accuracy toward validation, uncertainty, efficiency and whether foundation models can become scientific infrastructure.”
Built through mentorship and students
Alam credits Rizwan Uddin, NPRE professor and department head, as one of the strongest supporters of his development as an independent researcher. “Dr. Rizwan gave me the freedom to think broadly and take intellectual risks when these ideas were still forming,” Alam said. “I am deeply grateful for his mentorship and trust.”
He also credits Professor Souvik Chakraborty of IIT Delhi and a Fulbright Visiting Professor at UIUC for scientific collaboration in operator learning and scientific machine learning.
Doctoral researcher Kazuma Kobayashi's work on neural operators, digital twins and virtual sensing helped establish the technical foundation from which the cross-component foundation-model idea emerged.
“The award carries my name, but the research direction was built collectively,” Alam said. “My students have worked relentlessly on difficult ideas where none of us knew in advance whether they would succeed.”
Where can intelligence actually live?
Alam sees the CAREER project as part of a broader transition in AI: Smaller rather than simply larger. Sparse rather than dense. Hardware-aware. Physics-aware. Uncertainty-aware. Energy-aware.
“We have spent years asking how powerful AI can become,” Alam said. “One of the defining questions now is where that intelligence can actually live. For advanced reactors and other safety-critical systems, the answer cannot simply be a bigger model in a bigger data center.”
At the center of the project is a larger proposition: AI should not only become more intelligent. It should become deployable.