manufacturing · digital · materials · family: the answer is in a field you’ve never heard of
atomstoengineering,themiddleismissing
Linking materials models from atomic to engineering scales remains ambitious, and multiscale models of mechanical behavior have yet to reach industry
Problem statement
Predicting a material's engineering properties — strength, toughness, fatigue life, corrosion resistance — from its composition and processing requires linking models that work at very different length and time scales. The 2019 National Academies decadal survey describes modeling at quantum mechanical, atomic, mesoscale (coarse-grained or phase-field), and continuum scales: density functional theory (DFT) at the quantum level, molecular dynamics (MD) at the atomic scale, CALPHAD and phase-field methods at the mesoscale, and continuum process models used in manufacturing. Each has improved over the past decade, the survey says, but the links between them are unfinished: "Progress has been made in physics-based multiscale models for mechanical behavior prediction, but these have yet to be adopted by industry." The 2021 Materials Genome Initiative strategic plan calls linking knowledge from first-principles quantum mechanical calculations through the entire life cycle of a product "ambitious," requiring "multiple codes at various length and temporal scales." The survey places much of what matters in the middle: the atomic and mesoscale processes of phase transformation and of the formation and evolution of defects and interfaces "govern the synthesis and performance" of metals and of materials in general.
Why this matters
The Materials Genome Initiative (MGI), launched in 2011, set out "to discover, manufacture, and deploy advanced materials twice as fast, at a fraction of the cost." The decadal survey puts the typical time from discovery to deployment of a new material at 20 years for many industries. The initiative's approach is to use data and computational tools in concert with experiment. Where the pieces have been linked, development has sped up: the survey reports that Ford Motor Company used the Integrated Computational Materials Engineering (ICME) framework to cut the product development time of cast aluminum power train products by 15-25 percent, and that the time from development to deployment fell from 8.5 years for the steel Ferrium S53 to 4 years for Ferrium M54. But the survey also says that extensive and accessible databases exist for only a few material systems, and the 2021 strategic plan says that "there is still a great deal of work to be done to achieve the desired levels of integration and utility" of tools and data across materials research and development.
What’s been tried and why it hasn’t worked
The decadal survey describes intense effort on single- and multiscale methods that may be concurrent, hierarchical, or hybrid, and reports real gains at each scale along with limits. DFT packages are now readily available and calculate material properties with high fidelity. MD follows atomic movements only for very short periods and relies on force fields generated by DFT or by more approximate methods; the survey notes that for classical MD, as for time-dependent modeling generally, computational demands of time and memory rise rapidly as system size increases. Machine-learned interatomic potentials are now in use (the survey names artificial neural network potentials and Gaussian approximation potentials, and describes the neural-network representation of DFT potential energy surfaces as several orders of magnitude faster than DFT), and the survey says potentials with DFT-level accuracy that are efficient enough for realistic, laboratory-level length and time scales are still needed. At the mesoscale, CALPHAD predicts phase diagrams and thermodynamic behavior and is heavily supported by industry, while the transition of phase-field and kinetic Monte Carlo tools to industry is still in progress. The ICME approach seeks to integrate materials models across length scales to capture the relationships between synthesis and processing, structure, properties, and performance, but the survey notes that its initial successes depended on a wealth of data on specific materials systems generated over prior decades of research. The survey also warns that the MGI has tended over time to become an interaction between modeling, computing, and data communications, "leaving experiment behind," with vast databases of purely computational results that lack experimental validation.
What would unlock progress
The decadal survey's key recommendation in this area is that the U.S. government support the development of new computational and advanced data-analytic methods, new experimental tools to probe the properties of materials, and novel synthesis and processing methods. Three needs named in the sources bear directly on the multiscale problem. First, simple atomistic potentials with DFT-level accuracy that are efficient enough to simulate realistic, laboratory-level length and time scales; the survey says machine learning has helped develop such potentials. Second, machine learning models that, once trained, calculate a wide range of properties at speeds orders of magnitude faster than conventional computational methods. Third, experimental data to test the models against: the survey describes growth in 3D and 4D characterization, including in situ methods, "specifically geared toward quantifying mesoscale microstructure and response under stimuli" (X-ray tomography, diffraction-based X-ray microscopy, serial sectioning), and sees opportunities both in modeling at multiple length scales and in in situ characterization. The MGI strategic plan says advances in experimental tools are needed to generate the vast quantities of experimental data required to validate and refine predictions from theory, modeling, and simulation.
Entry points for student teams
A student team could implement a machine learning surrogate model that predicts grain size and phase fraction evolution during heat treatment of a well-studied alloy system (e.g., Ti-6Al-4V or AA7075) using published CALPHAD and phase-field simulation data, then validate against experimental literature. Alternatively, a team could benchmark the accuracy of different machine-learned interatomic potentials (MACE, NequIP, M3GNet) for predicting mechanical properties of a specific material, comparing predictions against experimental data to quantify the current accuracy gap. Neither project needs a special facility: the decadal survey itself cites the LAMMPS molecular dynamics code (shareware, in the survey's word), the NIST Interatomic Potentials Repository (https://www.ctcms.nist.gov/potentials/), the OpenKIM knowledgebase of interatomic models (https://openkim.org), and Open Calphad, which the survey describes as a free development. Relevant disciplines: materials science, computational physics, machine learning, mechanical engineering.
Genes
Structural cousins: stuck the same way, in other fields
Sources
"Frontiers of Materials Research: A Decadal Survey," National Academies of Sciences, Engineering, and Medicine, 2019. accessed 2026-02-16. Also: "Materials Genome Initiative Strategic Plan," Subcommittee on the Materials Genome Initiative, Committee on Technology, National Science and Technology Council, November 2021; "Integrated Computational Materials Engineering (ICME): Implementing ICME in the Aerospace, Automotive, and Maritime Industries," The Minerals, Metals, & Materials Society (TMS). go to source ↗
Working notes
Notes made while writing and checking this brief, shown as written. They use working shorthand.
- The Materials Genome Initiative was announced in 2011 by President Barack Obama; the 2019 NASEM decadal survey names it, with the National Nanotechnology Initiative, as one of two major government initiatives that stimulated U.S. materials research over the decade. Neither the survey nor the 2021 strategic plan singles out multiscale modeling as the field's central unsolved problem; the survey says physics-based multiscale models for mechanical behavior prediction have yet to be adopted by industry, and the plan calls the linking of first-principles calculations to the product life cycle ambitious.
- The `failure:disciplinary-silo` tag reflects this write-up's reading that the methods for each scale (quantum mechanical, atomic, mesoscale, continuum) are developed and used largely separately. The sources do not describe separate communities with different codes, validation standards, and publication venues; the decadal survey reports more multidisciplinary teams forming, and the strategic plan sets out to incentivize interdisciplinary teaming.
- The `temporal:newly-tractable` tag reflects the recent emergence of machine learning interatomic potentials and AI-driven materials discovery tools that offer new paths to bridge scales computationally.
- Cross-domain connection: shares the multi-scale prediction challenge with environment-critical-zone-process-prediction (predicting landscape-scale behavior from pore-scale and hillslope-scale processes) and bio-genotype-phenotype-prediction-gap (predicting organism-level properties from molecular-level information).
- The decadal survey reports industrial successes for the ICME approach (Ford's cast aluminum power train products; the Ferrium steels designed by QuesTek) and says many other companies, such as Boeing, use integrated approaches. It ties ICME's initial successes to data on specific materials systems built up over prior decades. The TMS implementation study's page describes the building blocks it sets out as written for the automotive, aerospace, and maritime sectors "but portable to other industries as well."
Reconciliation 2026-10-01: Fact repair after reading the whole write-up against the three sources it names, all fetched 2026-10-01. (a) The decadal survey, read in full (front matter, summary, chapters 1 to 5, appendices) in the National Academies online reader at https://www.nationalacademies.org/read/25244, which is where https://doi.org/10.17226/25244 leads. (b) The Materials Genome Initiative Strategic Plan, read in full; the original address on mgi.gov now returns "not found," so the Internet Archive copy was used: web.archive.org/web/20241127180215/www.mgi.gov/sites/default/files/documents/MGI-2021-Strategic-Plan.pdf. (c) The ICME implementation study: only its landing page could be read, from the Internet Archive copy web.archive.org/web/20230507135342/www.tms.org/ICMEStudy (tms.org blocks automated access and the report itself sits behind a download form), so the report's contents were not read. Changes: (1) The long title said materials science "cannot predict engineering properties from atomic structure because the mesoscale remains unsimulable." The survey says mesoscale modeling "has also had significant advancements over the past decade" and describes steels designed with computational methods and deployed; what it says is unfinished is that physics-based multiscale models for mechanical behavior prediction "have yet to be adopted by industry." Old title: "Materials Science Cannot Predict Engineering Properties from Atomic Structure Because the Mesoscale Remains Unsimulable." New title: "Linking Materials Models from Atomic to Engineering Scales Remains Ambitious, and Multiscale Models of Mechanical Behavior Have Yet to Reach Industry." The display title is unchanged. (2) The Source line gave the strategic plan's author as OSTP; its title page names the Subcommittee on the Materials Genome Initiative, Committee on Technology, National Science and Technology Council, November 2021. The Source line gave the third source as a 2018 NASEM report; the study is published by The Minerals, Metals, & Materials Society (TMS), and its page gives no 2018 date (the related items listed on it are dated 2013). The study's title is kept as it was, "Integrated Computational Materials Engineering (ICME): Implementing ICME in the Aerospace, Automotive, and Maritime Industries," which is the full title printed on the study flyer linked from the TMS page; the attribution now reads TMS, and no year is given because neither the page nor the flyer states one. (3) Problem Statement: "10 orders of magnitude," the scale figures (~0.1 nm, ~0.01-100 μm, ~mm-m), finite element analysis, "no reliable method bridges these scales," and the mesoscale as "the fundamental bottleneck" are not in any of the sources. The paragraph now gives the survey's own list of scales and methods, its sentence on multiscale models and industry, and the strategic plan's statement that linking first-principles calculations through a product's life cycle is ambitious. (4) Why This Matters: the MGI goal was given as halving time and cost; the survey quotes it as "twice as fast, at a fraction of the cost." "10-20 years and hundreds of millions of dollars" is now the survey's "typical 20 years for many industries"; no dollar figure is in the sources. Removed because none of the sources contains them: "$500M+ in federal investment"; that the 2021 plan "acknowledged that multi-scale modeling remains the field's central unsolved challenge" (the plan uses the word multiscale once, about data); the jet-engine, nuclear-reactor and vehicle-alloy examples; and the estimate attributed to Citrine Informatics and Materials Design that prediction could speed development 5-10 times (the survey mentions Citrine only as a data-services company). The paragraph now uses the survey's examples: Ford's 15-25 percent reduction in product development time for cast aluminum power train products, and 8.5 years for Ferrium S53 against 4 years for Ferrium M54. (5) What's Been Tried: not in the sources and removed: information loss at each handoff; DFT limited to ~100-1,000 atoms; MD up to ~10⁹ atoms; MACE and NequIP (both newer than the survey); machine-learned potentials "validated only against DFT benchmarks"; uncertain phase-field input parameters; CALPHAD covering "only equilibrium properties"; and ICME workflows being "bespoke" and not generalizing. The paragraph now reports what the survey says about each method, including its warning that the MGI has tended to leave experiment behind. (6) What Would Unlock Progress: the target volumes (~mm³) and timescales (seconds to hours), "surrogate models," and electron backscatter diffraction as an in-situ technique are not in the sources. The section now gives the survey's key recommendation, its stated need for efficient DFT-accuracy potentials, its account of machine learning speed, its description of 3D and 4D characterization, and the plan's statement on experimental data for validation. (7) Entry points: both projects kept as written; one sentence added naming the open tools the survey itself cites (LAMMPS, the NIST Interatomic Potentials Repository, OpenKIM, Open Calphad). The NIST and OpenKIM addresses answered on 2026-10-01. (8) Source Notes: "launched by OSTP" and "the most prominent federal effort" are now the survey's wording (announced by President Obama; one of two major government initiatives); the statement that both documents identified multiscale modeling as the central unsolved problem was replaced with what they do say. The sentence that Boeing, GE and Ford "have published case studies" is not in the sources; the note now gives the survey's Ford, QuesTek and Boeing statements and the TMS page's description of the building blocks it sets out. The note on the `failure:disciplinary-silo` tag now says plainly that the sources do not describe separate communities with different codes, validation standards and publication venues, and that the survey reports more multidisciplinary teams forming. Confirmed in the sources and left as written: the MGI's 2011 launch; the decadal survey's title, publisher, year and DOI (checked against the Crossref record and the reader's citation line); DFT, molecular dynamics, CALPHAD and phase-field as the methods at their respective scales; machine-learned interatomic potentials as a recent development (the basis of the `temporal:newly-tractable` note). Could not be checked: anything in the TMS study beyond its landing page; and, in the entry points, that published CALPHAD and phase-field data exist for Ti-6Al-4V or AA7075 and that MACE, NequIP and M3GNet are available for benchmarking, because no source named on this page covers them. Genome tags untouched. A second reader went over these repairs the same day, 2026-10-01, with the three sources fetched independently (plus the study flyer linked from the TMS page), and made four wording corrections: the TMS study's original title was kept, as described in (2); the sentence on DFT no longer says that time and memory limit how long a large system can be simulated, because the survey makes that statement about classical molecular dynamics and time-dependent modeling, not about DFT; the "several orders of magnitude faster than DFT" figure is now tied to the neural-network representation of DFT potential energy surfaces, as the survey has it, rather than to all machine-learned potentials; and the TMS page's "portable to other industries as well" is now attributed to the building blocks the study sets out, as the page has it, rather than to its frameworks. Every other repair stands.