Why do things break? People have been interested in controlling mechanical properties since time immemorial.
Understanding the connection between a material’s structure and its properties should enable intelligent design of materials with desired properties,
but it is difficult to navigate the vast complexity of structure-property space.
Nothing theoretically prevents us from plucking perfect targets out of design space,
but many strategies for materials design run up against practical limits of time and resources.
As a result, much of the exciting potential for areas like nanotechnology, bioinspired engineering, and hierarchical architectures remains unfulfilled.
In my thesis, I leverage bioinspired experimentation, physics-based molecular simulation, and machine learning models
to clarify structure-property space and achieve generative design across a variety of materials and properties of interest.
Controlled fracture paths, bioinspired crystals with tailored hardness values, and hierarchical honeycombs with dictated stress strain behavior
are only the beginnings of what modern AI-augmented research paradigms can accomplish.
Read the full thesis here
Can we reliably craft nanomaterials with controlled structure down to the molecular level?
By taking inspiration from nature, we can leverage the work of millions of years of evolution
as a starting point for design. For instance, biological cells have achieved incredibly complex
nanostructures via self-assembly of amphiphilic molecules. We can use similar strategies
to craft our own nanomaterials, but such amphiphile assemblies are not known to be mechanically robust.
In my work, I synthesized a series of novel amphiphiles incorporating molecular motifs from Kevlar,
a material commonly known for bulletproof vests, to enhance mechanical stability of assembled structures.
By varying molecular flexibility of Kevlar-based amphiphiles, I achieved a library
of self-assembled structures including nanoribbons, nanohelixes, and nanoplates.
I used the stiffest nanoribbon structures to fabricate centimeters-long hierarchically ordered threads
with unprecedented mechanical stability.
Ultimately, this patented Kevlar-based amphiphile platform serves as a fruitful basis for future nanomaterials design.
How can we probe the atomic mechanisms behind the most robust materials in nature?
While direct experimental manipulation of atoms remains difficult, computational physics-based
simulations can shed light on the dynamic behavior of molecules.
Here, I conduct such molecular dynamics simulations
to investigate the toughest materials
in biological systems - biominerals. Specifically, I investigate the problem of crystalline
fracture in biominerals such as aragonite, calcite, vaterite, and hydroxyapatite.
I demonstrate the ability to both identify a previously overlooked mechanism for mechanical
robustness and map out the applicability of a known mechanism in one biomineral to other biominerals.
Ultimately, understanding these mechanisms not only provide context into why we see certain patterns in nature,
but also serve as useful heuristics for designing similar bioinspired crystalline materials.
How do we explore and predict the unknown? Is it possible to pinpoint and instantiate the desired?
By applying machine learning models to the relationship between a material's structure and properties,
specifically strategies of surrogate modeling and latent space representation,
I reach predictive insights scalable in sample number, system size, and complexity.
Then, by considering the gestalt of bioinspiration, physics-based simulations, and machine learning,
I implement an end-to-end process for achieving generative design of material structures
with dictated mechanical properties.
I treat predictive crystalline material fracture, cantilever compliance, and notched beam buckling
by training Convolutional Neural Networks, Long-Short Term Memory networks, and Variational Autoencoders
on both experimental and computational datasets. In my generative design projects for obtaining target fracture behavior, hardness values, or stress-strain properties,
I additionally use Deep Residual Neural Networks, Generative Adversarial Networks, and Genetic Algorithms.
Ultimately, this cross-disciplinary paradigm allows us to reach farther than
any singular perspective in isolation, and the results here represent only the beginning of what is now possible.
