| dc.description.abstract | Humans depend on energy-absorbing structures constantly during daily life. Crumple zones in cars protect occupants during a crash. Packaging protects sensitive goods during delivery. Sports equipment protects athletes during both expected and unexpected impacts. Unfortunately, the development of energy absorbing structures is slow because phenomena such as high strain and self-self contacts make their performance difficult to model. Thus, physical experiments are often the only way to effectively evaluate the performance of a potential structure.
To tackle this problem, we employ a self-driving lab comprised of an automated experimentation system guided by machine learning. This system can in principle provide the acceleration in terms of experimental throughput and information per experiment needed to address the challenge of discovering high-performing structures. Initially, we developed a family of structures called generalized cylindrical shells that are reliable to additively manufacture and that include over trillions of unique designs. Next, we modified a self-driving lab and physically tested tens of thousands of structures in quasistatic compression. During a campaign lasting nearly two years, we discovered a component with an energy absorbing efficiency of 75.2%, the highest ever reported. Furthermore, using seven different polymer filaments, we found designs that outperformed all prior synthetic structures in energy absorbing efficiency across a broad stress range from 100 Pa to 10 MPa. In addition to this technical outcome, because this campaign lasted orders of magnitude longer than previously published self-driving lab campaigns, we also uncovered insights about the interactions between researchers and self-driving labs, such as how to monitor the progress of the campaign and how to adjust the campaign when problems arise.
Accounting for impact events further complicate the task of developing high-performing structures because of strain-rate dependent material behavior. We developed a physics-informed model to predict optimal impact velocity from a single quasistatic test. We then leveraged limited intermediate strain rate and impact testing to refine the model for strain-rate strengthening and demonstrate the model’s extrapolative abilities by applying it to different impactor masses, different designs, and a different material. This model can be used to screen potential designs using our extensive collection of quasistatic experiments or to guide future impact testing.
Finally, we explored the use of a self-driving lab as a community resource by collaborating with outside researchers to study the processing-structure-property relationships of foam-like structures printed on fused filament 3D printers using viscous thread printing. We ran two independent active learning campaigns that tested hundreds of samples, allowing us to develop models that predict their layer height, modulus, and stress-strain curve. These models not only allowed the realization of printed foams that match the stress-strain response of commercial foams, but it also enabled inverse design of 3D mechanisms by allowing designers to spatially modulate processing parameters. This work not only furthers the mechanical understanding of viscous thread printing, but also serves as a template for increasing the utilization of existing self-driving labs through collaboration. | en_US |
