Mapping Microstructure: Manifold Construction for Accelerated Materials Exploration
Simon A. Mason, Megna N. Shah, Jeffrey P. Simmons, Dennis M. Dimiduk, Stephen R. Niezgoda · 2025
Read Mapping Microstructure: Manifold Construction for Accelerated Materials Exploration by Simon A. Mason, Megna N. Shah, Jeffrey P. Simmons, Dennis M. Dimiduk, Stephen R. Niezgoda (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-15022.txt.
Introduction
Simon Mason and four co-authors map the inner structure of materials onto a low-dimensional surface in a 2025 paper. Faster work on new materials needs clear links from how a material is made to its inner structure and its traits. Their frame treats structure not as one image but as a spread of outcomes that one process can give.
The team assumes those outcomes lie on a low-dimensional surface set by a few key inputs. Runs of a phase-field model, where a mix splits into woven phases, give the test case. Three kinds of descriptors are scored: pair statistics, chord-length spreads, and shape tools from topology. Scores cover the true size of the latent space and how well process settings can be read back from it.
Catalog record
| CREATOR | Simon A. Mason, Megna N. Shah, Jeffrey P. Simmons, Dennis M. Dimiduk, Stephen R. Niezgoda |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-15022
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mapping Microstructure: Manifold Construction for Accelerated Materials Exploration Simon A. Mason, Megna N. Shah, Jeffrey P. Simmons, Dennis M. Dimiduk, Stephen R. Niezgoda Abstract Accelerating materials development requires quantitative linkages between processing, microstructure, and properties. In this work, we introduce a framework for mapping microstructure onto a low-dimensional material manifold that is parametrized by processing conditions. A key innovation is treating microstructure as a stochastic process, defined as a distribution of microstructural instances rather than a single image, enabling the extraction of material state descriptors that capture the essential process-dependent features. We leverage the manifold hypothesis to assert that microstructural outcomes lie on a low-dimensional latent space controlled by only a few parameters. Using phase-field simulations of spinodal decomposition as a model material system, we compare multiple microstructure descriptors (two-point statistics, chord-length distributions, and persistent homology) in terms of two criteria: (1) intrinsic dimensionality of the latent space, and (2) invertibility of the processing-to-structure mapping. The results demonstrate that distribution-based descriptors can recover a two-dimensional latent structure aligned with the true processing parameters, yielding an invertible and physically interpretable mapping between processing and microstructure. In contrast, descriptors that do not account for microstructure variability either overestimate dimensionality or lose predictive fidelity. The constructed material manifold is shown to be locally continuous, wherein small changes in process variables correspond to smooth changes in microstructure descriptors. This data-driven manifold mapping approach provides a quantitative foundation for microstructure-informed process design and paves the way toward closed-loop optimization of processing–structure–property relationships in an integrated materials engineering context. 1 Introduction The development of new materials is a fundamentally iterative process, with early metallurgical practices revolving around ad-hoc methods like “heat it and beat it” or “cook and look” to gradually refine manufacturing techniques. It is only in the last century or so with the advent of microscopy that metallurgy has transitioned from a craft discipline into materials science [ 1 ] . Much of the challenge in modern materials development results from the complex relationships between processing conditions and resultant properties, and slow exploration leaves new materials systems at risk of stagnating and falling into the developmental “Valley of Death” [ 2 ] . In an attempt to accelerate the development process, integrated computational materials engineering (ICME) methodologies like microstructure sensitive design (MSD) [ 3 ] and microstructure informatics (MI) [ 4 ] are being developed to establish data-driven approaches to the mapping of the processing-structure-property relationships (PSPRs). A key strategy for successful technology development as proposed by the National Research Council is through “quick, iterative development cycles”, with quantifiable measurements providing feedback information on the impact of developmental changes [ 2 ] . In order to enable these rapid cycles, data-driven practices and quantitative analysis can be used to formalize the relationships between processing, structure, properties, and performance, defining a mapping between these domains — termed the material manifold . For this purpose, we define the notion of navigability on the material manifold as the ability to make predictable changes between the processing and structure domains within a full feedback control loop. This enables iterative development and optimization to reach a target material structure having desired properties and performance, intuitively following the approach of a process design engineer. Stochastic descriptors permit knowledge of quantitative distances between material states, a necessary aspect of navigability. This work utilizes two concepts for establishing mappings between the processing and structure domains to enable informative, quantitative feedback for rapid material development: material state descriptors and the manifold hypothesis. The stochastic representation of microstructure allows us to establish material state descriptors by defining “microstructure” as a random variable controlled by processing conditions, rather than the more traditional interpretation of “microstructure” as an image [ 5 ] . Paired with the manifold hypothesis, which asserts that high-dimensional data can be “described by a function of only a few underlying parameters” [ 6 ] , the relationship between processing conditions and potential microstructure states can be mapped out across the design space, constructing the material manifold. We hypothesize that the stochastic representation of microstructure allows for the construction of a low-dimensional manifold, mapping the relationship between material processing conditions and resulting microstructure. We will test this hypothesis by comparing the dimensionality and invertibility of different descriptor methods to determine the stochastic representation’s ability to capture the controlling features of the underlying process. 2 Background 2.1 Microstructure as a Stochastic Process A foundational concept for understanding microstructure’s role in PSPRs is treating microstructure as a stochastic process [ 5 ] . In this definition, microstructure is a probability distribution of “essentially equivalent” microstructure realizations sampled from the stochastic process. This concept is generalized in this work to instead define the material state as a stochastic process - with microstructure being one descriptor set to represent material state information. The stochastic process is referred to as ‘big M ’ or the ‘ M -state’, and individual realizations within one M -state are known as ‘little m ’s’ or ‘ m -instances’. This M/m representation of material states and microstructure can be thought of as the difference in ‘microstructure as an image’ vs. ‘microstructure as a description of the state of a material,’ with the former being the m -instances and the latter being the M -state. While the term “microstructure” is multifaceted and can take on many meanings, it is often used to refer to an image, or a quantitative descriptor set derived from an image. From here, we use representative images and the term “microstructure” to mean a m -instance of a M -material state, or alternatively and interchangeably to mean the set of m -instances used to describe a M -state when used in that context. The M/m concept is not novel to this work, however it has mainly been used to describe distributions of m -instances within one M -state or to compare discrete classifications of M -states [ 5 ] . This work will expand it to allow for comparisons of a continuum of M -states across the material manifold. Figure 1 : Microstructure as a stochastic process. 2.2 Processing-Microstructure-Property Relationship and the Need for Multi-scale Representations The processing-structure-property relationship is a fundamental concept in materials science. Being able to harness the linkage between material processing, its impact on the resulting microstructure, and the property response of these materials, is a powerful tool for MSD [ 3 ] . Under this methodology, the material domain, design space, and resultant properties are explored with respect to the microstructure. The relationship between processing and properties is difficult to determine directly but can be understood by relating processing to microstructure and microstructure to properties [ 7 ] , as shown in Figure 2 . The processing-structure-property relationship requires a complete description from the microstructure to be able to understand how properties relate to processing conditions. Figure 2 : Processing-structure-property relationship triangle. As of today, this “microstructure bridge” is a primarily academic construct, but increasing complexities in the processing domains and material performance requirements necessitate the development of tools to explore their relationships to microstructure. In practical engineering and historical contexts, materials were explored, developed, and defined by their properties with no insight into microstructure, either because the characterization tools did not exist, or