IRG1: Trainable Soft Materials

Senior Participants (*=coordinators): A. Dinner, M. Driscoll, A. Esser-Kahn, *H. Jaeger, A. Murugan, S. Nagel, M. Olvera de la Cruz, S. Palmer, *S. Rowan, B. Tian, T. Witten

Can we train a material to exhibit desired properties, and then retrain it to exhibit different behavior? Materials design focuses on establishing specific structural configurations and interactions among the constituent components of a material, often at molecular scales. Once identified, the associated design parameters typically are intended to remain fixed, so as to maintain a material’s properties. Changing the targeted properties then requires careful parameter re-programming. It is fascinating how biology takes a different route and generates functionality by allowing the design parameters to evolve and adapt as the system is affected by the environment. An example is bone, which is a material that reconfigures to become stronger or more resilient at precisely those places where repeated stressing provides the cues. IRG 1 will take this lesson from the biological rules of life to mimic the adaptation that occurs in biological systems in physical materials with a goal of creating novel functionality through the process of training. Since the adaptation during training is carried out by the material itself, it can be achieved without precise (re-)design of the local structure. We will also bring understanding of what features of the intricate dance between a biological structure and its environment are useful to the material world. In return, as we understand more about trainability of materials, it is expected that this can help materials biology and advance mastery of synthetic biological systems by identifying training mechanisms at play in the biological realm.

Materials training falls under the larger umbrella of materials processing but, as defined here through the link with biology, it differs in several key aspects from standard processing methods and opens up a range of new opportunities. In particular, while training is applied at length scales much larger than individual constituent units, for example by manipulating the material as whole, intrinsic material properties are exploited to direct the training action to only those constituent parts that need to be evolved, but not others. In other words, during training the material itself ‘decides’ where and how to evolve internally. This mimics what happens during athletic or musical training (i.e. different training protocols are used to train different muscles in the body) or in the case of bone mentioned above. Taking this a step further, we envision a situation where different material properties, or functions, emerge simply from changing the training protocol (see below). Beyond this, training can achieve what can be called ‘material learning’: the material can be trained to exhibit specific, desired responses to different external stimuli. We envision adaptive impact-absorbing materials that can learn to redirect internal stresses by evolving their ratio of bulk to shear modulus, shape morphing materials that learn to self-fold into different configurations, or materials that work as actuators with multiple, trainable functionalities by learning different spatial or temporal patterns of applied stress and responding with pattern-specific outcomes. In all these cases, re-training using a different protocol of applying mechanical stimuli might be used to modify the material to generate a new response. This highlights a difference with standard materials processing methods, which condition a material for a single purpose.

Soft materials are particularly suited for training as they often have a multitude of easily accessible, energetically similar configurations that training can select so as to amplify a desired property. In soft materials, training protocols based on applying mechanical stress provide straightforward access to nonlinear regimes and allow for the activation of chemical pathways, both of which can help to imprint a memory by triggering long-lasting changes such as plastic structural deformation. Many materials and patterns can be profitably modeled as networks. Macroscopic mechanical metamaterials composed of nodes and struts, crosslinked polymers and biological fibers, the abovementioned bone, and even the creases in folded sheets all have a network structure where links between nodes can be clearly identified. We can then think of training as evolving the properties of the links and/or the configuration of the nodes. Thus, viewing them as prototypical adaptive materials we focus on networks for addressing the goals of this IRG. Specific types of soft matter networks that share similar characteristics will be addressed in the three Focus Areas (FAs) presented below.

Clearly, not all materials can be trained toward a useful outcome and proper choice of a training protocol will be required to reach a desired material performance. Indeed, a key goal of the IRG is to understand what is required in a material for it to be ‘trainable’/’retrainable’ and what are the possibilities and limits of training. Trainability, together with learning and memory, forms a key element in the evolution of biological systems and, more recently, in machine learning with computers. Aspects of training have also been used in materials processing — to harden metal, align polymer domains, or activate current paths in neuromorphic computation devices. However, there currently is no systematic framework for designing trainable soft materials and for establishing their training protocols. This IRG aims to develop such a framework by taking advantage of an emerging synergy between recent ideas from biology, materials science, polymer chemistry and soft matter physics, all addressing different aspects of learning and memory in complex systems. We build on pioneering initial research at the UChicago MRSEC, including a recent iSuperSeed2 effort aligned with Understanding the Rules of Life. This iSuperSeed2 indicated the exciting potential of a training-based approach for creating multifunctional, pluripotent materials. At the same time, it showed that insights gained can aid the understanding of biological systems in revealing principles of adaptability.

The new types of trainable materials we envision can exist from the macroscopic to the molecular scale; in their most potent manifestations they will require careful consideration of all of these scales. To achieve this vision, IRG 1 brings together expertise in polymer synthesis/characterization (Esser-Kahn, Rowan), nano-to-macro experiments (Driscoll, Jaeger, Nagel, Tian), soft-materials modeling (Olvera de la Cruz, Witten), bio-soft matter theory (Dinner, Murugan), and learning in neural networks (Palmer). FAs within the IRG explore different aspects of training in three classes of soft matter: Macroscopic network-based materials (FA1), Dynamic polymer networks (FA2), and Particle/gel-network composites (FA3). FA1 focuses on training by exploiting adaptivity of the links, while in FA2 and FA3 the configuration of the nodes can also change during training. All three FAs address similar key questions concerning trainability, but explore them within the context of specific material systems. In particular, we ask how training can select specific, targeted behaviors or properties from a large number of possible outcomes, how many different, input-specific responses can be learned, and how suitable training protocols can be identified. Our goal of this concerted effort is to investigate the scope, limits and benefits of training as a new paradigm for generating novel multi-functionality in soft materials.

Traditional materials design vs. a materials training approach and the three Focus Areas (FAs) of the IRG.