Inverse design of soft materials via a deep learning–based evolutionary strategy

Abstract

Colloidal self-assembly—the spontaneous organization of colloids into ordered structures—has been considered key to produce next-generation materials. However, the present-day staggering variety of colloidal building blocks and the limitless number of thermodynamic conditions make a systematic exploration intractable. The true challenge in this field is to turn this logic around and to develop a robust, versatile algorithm to inverse design colloids that self-assemble into a target structure. Here, we introduce a generic inverse design method to efficiently reverse-engineer crystals, quasicrystals, and liquid crystals by targeting their diffraction patterns. Our algorithm relies on the synergetic use of an evolutionary strategy for parameter optimization, and a convolutional neural network as an order parameter, and provides a way forward for the inverse design of experimentally feasible colloidal interactions, specifically optimized to stabilize the desired structure.

Machine-learning approach for discovery of conventional superconductors

“…Finally, the ML models will be integrated in an inverse design strategy to explore the practically infinite materials space in an efficient manner. Currently, (inverse) design of functional materials with targeted properties is a very active research area with many success stories [94][95][96][97][98][99][100][101]. We hope that superconducting materials discoveries can be added to this list in the near future.…”

Predicting stress, strain and deformation fields in materials and structures with graph neural networks

“…Emergence and growth of the machine learning (ML) field in the recent years has though demonstrated the possibility to outperform traditional numerical solvers, greatly speeding up simulations of physical systems 17 – 22 , from the use of physics-informed neural networks to extract velocity and pressure fields from flow visualization 23 to the inverse-design of architected materials with extreme elastic properties using generative adversarial networks 24 . Given the importance of materials discover and design, linking materials’ micro- and meso- structure to mechanical properties (structure-to-property) 25 – 30 and inverse-designing (i.e., given targeted properties, finding optimal designs) high-performing architected metamaterials 10 , 13 , 24 , 31 – 39 have recently dominated the research scene. In both cases, materials performance is essentially dictated by local mechanical fields, such as stress and strain distributions, because of the effect of geometry, base materials’ behavior, and boundary conditions.…”

Learning to learn: Non-equilibrium design protocols for adaptable materials

“…This success can be attributed to "meta-properties" like modularity [1][2][3][4][5], robustness [6], plasticity for learning [7], and multifunctionality [8][9][10][11]. While inverse materials design has sought to optimize specific properties [12][13][14][15][16][17][18][19][20], less attention has been given to identifying general design strategies for creating materials with meta-properties.…”

“…Our method functions as a wrapper to existing optimization algorithms. It is therefore compatible with a wide range of pre-existing materials design procedures, ranging from fully computer-based [12,14] to fully in situ [26,27].…”