AI moves from specialty to infrastructure at MIT
Artificial intelligence has become a working layer across MIT research, according to an MIT Technology Review feature on how the institute’s labs are deploying the technology. The story describes AI not as a separate discipline confined to computer science, but as a tool now embedded in mechanical engineering, aerospace materials, energy systems and experimental science.
The shift is visible in the work of Sili Deng, an associate professor of mechanical engineering who studies combustion kinetics, emissions reduction and flame synthesis of energy materials. Deng’s path into AI was shaped by the disruption of the covid pandemic. After joining MIT’s faculty in 2019, she was setting up her lab when renovations halted. Rather than wait, she asked her team to test where machine learning could fill gaps in their combustion research.
Digital twins for combustion systems
Deng’s Energy and Nanotechnology Group used AI to develop a digital twin that mirrors the performance of an energy and flow device. A digital twin is a computational replica of a physical system, designed to reflect how that system behaves under changing conditions. In this case, the long-term aim is to predict and control fuel combustion systems in real time.
That target matters because combustion systems remain central to many parts of the energy and transportation economy. Better prediction and control could help researchers understand performance, emissions and operational stability in ways that are difficult to capture through physical testing alone. The source material does not claim the system has already reached full real-time control in production. It says the model should eventually be able to predict and control the workings of fuel combustion systems in real time.
The example shows one reason AI has spread quickly through research labs: it can extend existing scientific knowledge rather than replace it. Deng’s team approached machine learning through the lens of combustion fundamentals, asking where existing methods had gaps. That framing is different from treating AI as a general-purpose shortcut. It uses domain expertise to define useful questions and evaluate whether model output makes sense.
