Designing athermal disordered solids with automatic differentiation

TL;DR

This paper introduces a differentiable programming approach to design disordered solids with targeted properties, enabling inverse self-assembly without relying on a well-defined structure, and demonstrating control over multiple material properties.

Contribution

It presents a novel method using automatic differentiation to directly optimize interactions for desired properties in disordered solids, bridging the gap between crystal assembly and amorphous material design.

Findings

Successfully tuned Poisson's ratio in disordered solids

Simultaneously optimized multiple properties like pressure and structural order

Method is scalable, robust, and transferable across systems

Abstract

The ability to control forces between sub-micron-scale building blocks offers considerable potential for designing new materials through self-assembly. A typical paradigm is to first identify a particular (crystal) structure that has some desired property, and then design building-block interactions so that this structure assembles spontaneously. While significant theoretical and experimental progress has been made in assembling complicated structures in a variety of systems, this two-step paradigm fundamentally fails for structurally disordered solids, which lack a well-defined structure to use as a target. Here we show that disordered solids can still be treated from an inverse self-assembly perspective by targeting material properties directly. Using the Poisson's ratio, $\nu$, as a primary example, we show how differentiable programming connects experimentally relevant interaction parameters with emergent behavior, allowing us to iteratively "train" the system until we find the set of interactions that leads to the Poisson's ratio we desire. Beyond the Poisson's ratio, we also tune the pressure and a measure of local 8-fold structural order, as well as multiple properties simultaneously, demonstrating the potential for nontrivial design in disordered solids. This approach is highly robust, transferable, and scalable, can handle a wide variety of model systems, properties of interest, and preparation dynamics, and can optimize over 100s or even 1000s of parameters. This result connects the fields of disordered solids and inverse self-assembly, indicating that many of the tools and ideas that have been developed to understand the assembly of crystals can also be used to control the properties of disordered solids.