As an undergraduate majoring in robotics and intelligent systems at the University of Science and Technology of China, He Luan was introduced to a new but fast growing engineering field: 3D printing. “It gave me an idea of how revolutionary 3D printing could become in terms of how we make things and also what the challenges were to improving it,” she recalls. The insight spurred Luan to enroll in the Ph.D. program in industrial and systems engineering at the University of Southern California, where she specializes in applying data and statistical learning to 3D printing research and applies knowledge gained from her studies in USC’s computer science master’s degree program.
HP: What’s your HP Labs research project this summer?
A major initiative in the Print Adjacencies and 3D Lab explores ways to analyze job data and data from heat sensors in HP’s 3D printers to predict how well the printing process is working and then asks how we can use those predictions to optimize the process. As part of that effort, I’m applying deep learning to these data sets and exploring different neutral network architectures to predict the thermal behavior of each layer of material as it is printed. If we are successful, our goal is to use this model to make the machines print even more precisely than they can at present.
HP: How has it been going so far?
I’m in my sixth week of fourteen and I have an outline for what I hope to achieve and a plan for getting there. The data source is very rich and there’s deep science inside our printing physics, so it’s a challenge. We are actually designing our own neutral network architectures. We’ve done some preliminary tests. So far so good.
HP: What’s the value of using deep learning here?
We are trying to predict thermal behavior using models derived from physical sciences. The challenge is that thermal behavior in a 3D printer is incredibly complex. Other HP research teams are running a lot of experiments to help us to uncover the underlying interactions, and thanks to our collaborators in HP’s 3D printing business unit, we have great access to that data. Deep learning allows us to examine all the different data sources and automate the discovery process for the underlying inter-connectivity between different factors. By applying deep learning to the data, we’re hoping to reveal information patterns that allow us to predict thermal behavior and help us to build even more accurate physical models.
HP: What is the biggest challenge you face in doing this work?
Our problem is a bit different from a typical deep learning problem like video prediction, for example. In our case the predictions we come up with have to be very high resolution – at the scale of a prediction per voxel (the 3D equivalent of a pixel). That’s not always something you need in a typical deep learning application - when it can be enough just to be pointed in a better direction. Like most recurrent neutral network problems, we have both tempo and spatial components. Our system is different in that we have external spatial excitations in the form of agent amount data flowing to the printheads. To attack that, our “Big Idea” is to generate a new, scalable network structure. We may write an invention disclosure by the end of summer if we have good results.
HP: How does your HP work relate to your academic research?
My Ph.D. advisor has been collaborating with HP. In fact, he is visiting HP’s 3D printing group in Barcelona next month. My dissertation is looking at ways to improve geometric accuracy, or what we call geometrical fidelity control, in 3D printing. This is one of the leading problems that HP is interested in. I am definitely hoping to continue working with HP Labs after my internship. I’d love to help some of the ideas we cook up over the summer find their way into products.