Accelerating mechanical design using autonomous experimentation
Embargo Date
2022-09-26
OA Version
Citation
Abstract
The high level of control afforded by modern advancements in additive manufacturing (AM) has enabled unprecedent design freedom to explore novel designs and structures. This added complexity deepens the challenge of identifying optimized structures in a vast design space and introduces a host of new defects for which researchers and practitioners do not have the benefit of empirical engineering guidelines built upon decades of intense study. For several important nonlinear inelastic properties that are challenging to reliably simulate, physical experimentation remains the gold standard for investigation. As a result, the design process occurs manually through iterative manufacturing and testing. In this dissertation, we propose, evaluate, and utilize new approaches to accelerate the design process of mechanical structures that fuse advancements in AM, automation, and machine learning (ML).
First, we show the use of autonomous experimentation (AE) for removing the bottleneck created by the slow pace of manual iterative cycles in manufacturing and testing. Specifically, we developed a Bayesian experimental autonomous researcher (BEAR) for mechanical design that can plan and conduct its own experiments without human intervention by combining automated experimentation and active learning. The high-throughput of this system, relative to manual testing, allows for the comprehensive exploration of a large family of structures using thousands of experiments. We present the results of experimental campaigns conducted by the BEAR for optimizing toughness which resulted in the identification of high performing structures in 60 times fewer experiments than were required when using a grid-based campaign.
Next, we demonstrate that by using transfer learning, knowledge from simulation can be incorporated with the BEAR to accelerate the pace of mechanical optimization. Initially, we study a property that is well predicted by FEA, namely resilience, and find that knowledge from FEA can be transferred to the BEAR to reduce by ten-fold the number of experiments needed to find high performing structures. Additionally, we study toughness, a property not well predicted by FEA, and find that knowledge from FEA predictions of properties such as yield force can accelerate learning of toughness by transforming the experimental data and guiding experiment selection.
Finally, to investigate the utility of knowledge transfer between testing regimes, we use transfer learning to use data on quasi-static performance to design high-performing impact resistant structures. High impact resistance is a critical and desirable property that relies on the co-optimization of distinct properties ranging from compliance, strength, resilience, and toughness. Using feature representation transfer learning and data from both quasi-static and impact testing, we build a data-driven model that predicts impact performance based only on quasi-static testing, allowing for rapid exploration of novel designs.
Considering the multitude of design applications where experimental measurement of component properties is required for quantitative understanding, AE and transfer learning present unprecedented opportunities to efficiently explore vast design spaces that capitalize on the design freedom presented by AM. The principles described herein motivate future work in the further development and application of AE which could have a transformative impact in fields such as mechanics, materials, chemistry, and biology.