Rutgers-led research leads to autonomous smart systems that enable resilient 3D printing in space, war zones and disaster areas
Research led by Rutgers engineers has shown how artificial intelligence (AI) can solve two of the biggest challenges in manufacturing.
In two separate studies led by Rajiv Malhotra, an Associate Professor in the Department of Mechanical and Aerospace Engineering at the Rutgers School of Engineering, researchers have revealed how smart autonomous systems can make 3D printing reliable in extreme environments such as space and war zones, and how AI can significantly speed up innovation in manufacturing by reducing the need for costly experiments.
One study, published in The Journal of Manufacturing Processes, centers on what Malhotra calls "expeditionary additive manufacturing."
This refers to making parts, such as the interior or exterior of a spacecraft, on battlefields and in disaster zones which are outside the stable and controlled conditions of a factory. These environments are unpredictable and impose disturbances such as vibrations or temperature shifts that can ruin a print job. Operators in these settings, including astronauts or soldiers, may lack specialized training, which adds another layer of complexity.
"We were trying to understand how we can make expeditionary additive manufacturing robust to such unknown and disruptive disturbances," Malhotra said.
The solution members of the team devised could be lifesaving, he said. In space, a broken part could mean a failed mission. In a war zone, it could result in a vehicle that cannot move. In a disaster area, it could create a scenario where rescuers lack the tools needed to help people.
"Whether or not your part will turn out or not," Malhotra said, "can have missions fail completely. You could have people die."
To tackle this, Malhotra and collaborators developed a new AI technique called conditional reinforcement learning. The system uses a camera to monitor the printing process and instantly adjusts the printer's settings when it detects a defect. It doesn't need to know what went wrong ahead of time and can fix problems in just one step. It also does not need retraining to adjust to new conditions, thus overcoming a major challenge in such control techniques today.
"We created a tool which addresses that issue," Malhotra said. "We don't have to anticipate anymore. Whatever disturbances come, we can deal with it without throwing away the part or stopping failure, both of which are bad for mission assurance."
The system is designed to work without stopping the printer or retraining the software, which is crucial in emergencies. It also doesn't rely on the user being an expert. In fact, the AI treats lack of training as just another disturbance to overcome.
We trained the AI to expect the unexpected, rather than expect the expected.
Rajiv Malhotra
Associate Professor, Rutgers School of Engineering
"We trained the AI to expect the unexpected, rather than expect the expected," Malhotra said. "We have created a new AI technique that 'robustifies' expeditionary manufacturing beyond the reach of literature. It reduces defects by 10 times or more, increasing quality by similar amounts even when the disturbances are not known in advance."
Given that expeditionary manufacturing is critical for defense, space and disaster recovery applications, he said, "the resilience we achieved is both critical and hitherto unrealized in the state of the art."
In another study, published in The Journal of Intelligent Manufacturing, Malhotra and collaborators focused on speeding up innovation in conventional manufacturing.
Engineers have long relied on physics-based models and trial-and-error testing to understand how materials behave during production. These models can take decades to develop and require deep expertise. Even modern machine learning tools need massive datasets to work well.
"That process is very slow," Malhotra said. "It is very prone to error. It can take 30 years sometimes to really develop a process."
His team created an AI system that reads scientific papers, pulls out useful information and combines it with a small amount of experimental data to build predictive models. The AI performs as an expert would, learning from past research and refining its predictions with each iteration.
"The AI acts like that Ph.D. expert," he said. "It tries a few times, and then it gets it right."
Instead of running hundreds of experiments, the team achieved accurate results with just 30 samples. "We cut short the samples that you have to make," Malhotra said. "That means you're doing things much faster."
The system uses large language models to retrieve knowledge from scientific literature and refine it using real-world data. "Our job really is to take an existing AI system and not spend $5 billion, but spend really next to nothing to say, 'Find me a hypothesis that works for my case,'" Malhotra said.
This innovation could accelerate development in industries including aerospace, automotive, electronics, and defense. "This method reduces the need for human interpretation and large experiments, speeding up innovation for new or complex manufacturing processes," he said.
Both studies were collaborative efforts. On the expeditionary manufacturing paper, Malhotra worked with Jeremy Cleeman, a doctoral student, and Adrian Jackson, an undergraduate from the Rutgers School of Engineering. He also partnered with Shane Esola from the U.S. Army Armaments Graduate School, Chenhui Shao from the University of Michigan, and Hongyi Xu from the University of Connecticut. For the study involving large language models, Malhotra's collaborators included Kiarash Naghavi Khanghah and Hongyi Xu from the University of Connecticut, and AnandKumar Patel, a doctoral student from the Rutgers School of Engineering.
Explore more of the ways Rutgers research is shaping the future.
