Designing lattices for impact protection using transfer learning

• Published in: Matter Vol. 5; no. 9; pp. 2829 - 2846
• Main Authors: Gongora, Aldair E., Snapp, Kelsey L., Pang, Richard, Tiano, Thomas M., Reyes, Kristofer G., Whiting, Emily, Lawton, Timothy J., Morgan, Elise F., Brown, Keith A.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 07.09.2022
• Subjects: additive manufacturing; automated experimentation; impact protection; lattice design; machine learning; structured matter; transfer learning; impact protection; transfer learning; additive manufacturing; structured matter; automated experimentation; lattice design; machine learning; MAP3: Understanding
• ISSN: 2590-2385; 2590-2385
• DOI: 10.1016/j.matt.2022.06.051

Like many specialty applications, the pace of designing structures for impact protection is limited by its reliance on specialized testing. Here, we develop a transfer learning approach to determine how more widely available quasi-static testing can be used to predict impact protection. We first extensively test a parametric family of lattices in both impact and quasi-static domains and train a model that predicts impact performance to within 8% using only quasi-static measurements. Next, we test the transferability of this model using a distinct family of lattices and find that performance rank was well predicted even for structures whose behavior extrapolated beyond the training set. Finally, we combine 812 quasi-static and 141 impact tests to train a model that predicts absolute impact performance of novel lattices with 18% error. These results highlight a path for accelerating design for specialty applications and that transferrable mechanical insight can be obtained in a data-driven manner. [Display omitted] •Novel families of 3D-printed lattices evaluated using automated mechanical testing•Machine-learning model predicts impact performance using quasi-static behavior•Lattices with optimized impact performance achieved using a data-driven approach•Illustrated how specialized performance can be transfer learned from general testing Many facets of mechanical performance evaluation require using specialized experiments that are slow and expensive. For example, when designing components for impact conditions such as the crumple zone of a car, testing is expensive and destructive and requires large samples. In this study, we explore whether comparatively simple experiments can provide insight into highly specialized performance such as in impact. We fabricated lattices using 3D printing and tested them under quasi-static compression, which is more general and accommodates smaller samples than does impact testing. Using a database accumulated using automated quasi-static testing, we trained a machine-learning model that predicts impact performance from quasi-static data alone. In addition to accelerating the design of impact-resistant structures for safety applications, this work more generally illustrates how easy-to-acquire experimental data can allow one to better select samples for specialized experiments. We report a transfer learning approach for using general testing to predict the specialized performance of 3D-printed lattices. Specifically, we curated a large dataset of quasi-static tests using automated testing and learned a general featurization. Next, we connected this featurization to experimentally measured impact performance and found that the model predicted the impact performance of novel lattice families with 18% error. These results show how design for specialty applications can be accelerated in a data-driven manner.