Session: Rising Stars of Mechanical Engineering Celebration & Showcase

Paper Number: 148727

148727 - Smart Additive Manufacturing: M4am - Modeling, Materials Science, Monitoring, and Machine Learning for Rapid Part Quality Assurance 

What’s this research about, and who cares? 

The goal of this research is to realize the flaw-free, industrial-scale production of metal parts using additive manufacturing processes (metal AM/3D printing). Twenty pounds of raw material are currently required to make a one-pound part for the aerospace industry using subtractive machining. AM can reduce this so-called buy-to-fly ratio of 20:1 to 2:1, while simultaneously reducing lead time from six months to one week. Despite these advantages, industries are hesitant to adopt AM due to process inconsistency - parts may have undetected defects, such as porosity, that make them unsafe for use in mission-critical applications. It is estimated that trial-and-error part qualification requires upwards of $10 million and 3 years to complete.

What are the limitations with the existing approaches

With the advent of pervasive sensing, large data sets, and computing abilities, quality assurance of AM processes has entered the realm of Big Data. There is an instinct to use the sensor data with advanced deep learning AI methods for part quality assurance. For instance, terabytes of meltpool images have been acquired and analyzed using convolutional neural networks for predicting onset of porosity in laser powder bed fusion. Apart from concerns regarding storage and curation of the sensor data, the use of AI techniques has certain limitations. These limitations are: (1) need for large amounts of training and testing data, (2) limited transferability across different part shapes, and  (3) "the casualty of physics" - in that there is little consideration of the physics of the process. There are two possible alternatives. The first is to use a pragmatic AI methods, where instead of using raw sensor data directly into machine learning models, certain features, informed from understanding of the physics are extracted from the sensor data. The second approach, called gray-box or digital twins, is to combine thermal physics of the process via rapid process simulations with real-time data, and machine learning to predicit, detect, and correct flaws during the process. In this poster, I will make a case for pragmatic AI and digital twin methods through industry case studies.

What's new and unique about this work? 

This research seamlessly integrates the 4 Ms of AM to achieve rapid qualification of AM parts: (1) computational heat transfer modeling; (2) materials science to understand the link between thermal physics of the process and flaw formation; (3) monitoring (tracking) the process using in-situ sensor arrays; (4) machine learning (analytics) to detect and prevent flaw formation combining insights from physical models and sensor data. In this poster I present results from industry-based projects exemplifying the advantages of integrating fundamental heat transfer modeling with in-situ data for flaw mitigation, and predicting properties of AM parts. 

This research has produced two breakthroughs: (1) a novel mesh-less approach for rapid thermal modeling of AM processes, that enables design and parameter optimization to avoid flaw formation. The approach is 20x faster than existing finite element modeling; (2) combining physics-based simulations with real-time sensor data within computationally tractable machine learning models to predict multi-scale flaw formation ranging from microstructure evolved, surface finish, porosity, and geometric distortion.  

What's been achieved from an education and broader impacts perspective? 

(1) 6 Ph.D's graduated (3 ongoing, 1 minority), 2 MS, 6 REUs; 40+ peer-reviewed journal papers; 7 patent disclosures (1 granted);  20+ invited talks; and $2+ million in follow up funding; (2) Collaboration with Native American scholars from Navajo Technical University resulting in 4 joint papers; (3) 10 symposia organized at ASME and TMS (SFF) conferences; and (4) Ongoing commercialization of modeling and monitoring tools with a venture capital startup. 

Presenting Author: Prahalad Rao Virginia Tech

Presenting Author Biography: Prahalada Rao’s scholastic passion is captured in three words: Manufacturing, Sensing, and Analytics. Rao is an Associate Professor in Industrial Engineering at Virginia Tech. Rao has published over 70+ peer-reviewed papers. His research has garnered over $3 million in funding from federal agencies, including NSF (including CAREER award), DOE, NIST, and Office of Naval Research. He earned the 2017 Society of Manufacturing Engineers, Outstanding Young Manufacturing Engineer Award, and the 2019 University of Nebraska, College of Engineering Research and Creativity Award.

Authors:

Prahalad Rao Virginia Tech